Adapting a Pre-trained Single-Cell Foundation Model to Spatial Gene Expression Generation from Histology Images

Donghai Fang; Yongheng Li; Zhen Wang; Yuansong Zeng; Wenwen Min

cs.CVMar 20, 2026

Adapting a Pre-trained Single-Cell Foundation Model to Spatial Gene Expression Generation from Histology Images

Authors: Donghai Fang, Yongheng Li, Zhen Wang, Yuansong Zeng, Wenwen Min

Organizations: Sun Yat-sen University, China · Yunnan University, China · Chongqing University, China

Abstract

Spatial transcriptomics (ST) enables spot-level in situ expression profiling, but its high cost and limited throughput motivate predicting expression directly from HE-stained histology. Recent advances explore using score- or flow-based generative models to estimate the conditional distribution of gene expression from histology, offering a flexible alternative to deterministic regression approaches. However, most existing generative approaches omit explicit modeling of gene-gene dependencies, undermining biological coherence. Single-cell foundation models (sc-FMs), pre-trained across diverse cell populations, capture these critical gene relationships that histology alone cannot reveal. Yet, applying expression-only sc-FMs to histology-conditioned expression modeling is nontrivial due to the absence of a visual pathway, a mismatch between their pre-training and conditional ST objectives, and the scarcity of mixed-cell ST supervision. To address these challenges, we propose HINGE (HIstology-coNditioned GEneration), which retrofits a pre-trained sc-FM into a conditional expression generator while mostly preserving its learned gene relationships. We achieve this by introducing SoftAdaLN, a lightweight, identity-initialized modulation that injects layer-wise visual context into the backbone, coupled with an expression-space masked diffusion objective and a warm-start curriculum to ensure objective alignment and training stability. Evaluated on three ST datasets, ours outperforms state-of-the-art baselines on mean Pearson correlation and yields more accurate spatial marker expression patterns and higher pairwise co-expression consistency, establishing a practical route to adapt pre-trained sc-FMs for histology-conditioned spatial expression generation.

Figures & tables

Figure 1: Overview of HINGE. (a) Depicts the CellFM architecture, which is a single-cell foundation model (sc - FM) pre-trained on scRNA-seq with masked autoencoding. (b) In HINGE, the conditional denoising model is instantiated from CellFM and augmented with identity-initialized SoftAdaLN that injects histology and timestep context into each transformer layer within a stochastic masked diffusion process. This design keeps the training objective aligned with CellFM’s masked autoencoding for coherence in ST, thereby largely preserving the gene relationships learned from scRNA-seq.

Discriminative
Methods		cSCC				Her2ST				Kidney
Methods		PCC-50 $\uparrow$	PCC-200 $\uparrow$	MSE $\downarrow$	MAE $\downarrow$	PCC-50 $\uparrow$	PCC-200 $\uparrow$	MSE $\downarrow$	MAE $\downarrow$	PCC-50 $\uparrow$	PCC-200 $\uparrow$	MSE $\downarrow$	MAE $\downarrow$
	ST-Net [ 12 ]	$0.548_{\scriptstyle\pm.039}$	$0.448_{\scriptstyle\pm.043}$	$1.174_{\scriptstyle\pm.109}$	$0.803_{\scriptstyle\pm.041}$	$0.439_{\scriptstyle\pm.030}$	$0.323_{\scriptstyle\pm.031}$	$1.132_{\scriptstyle\pm.095}$	$0.837_{\scriptstyle\pm.039}$	$0.327_{\scriptstyle\pm.005}$	$0.209_{\scriptstyle\pm.005}$	$1.566_{\scriptstyle\pm.004}$	$0.982_{\scriptstyle\pm.002}$
	BLEEP [ 36 ]	$0.643_{\scriptstyle\pm.009}$	$0.548_{\scriptstyle\pm.012}$	$0.925_{\scriptstyle\pm.066}$	$0.719_{\scriptstyle\pm.035}$	$0.520_{\scriptstyle\pm.021}$	$0.400_{\scriptstyle\pm.018}$	$0.949_{\scriptstyle\pm.072}$	$0.754_{\scriptstyle\pm.029}$	$0.404_{\scriptstyle\pm.011}$	$0.285_{\scriptstyle\pm.010}$	$1.493_{\scriptstyle\pm.010}$	$0.965_{\scriptstyle\pm.003}$
	TRIPLEX [ 5 ]	$\underline{0.683_{\scriptstyle\pm.017}}$	$\underline{0.588_{\scriptstyle\pm.017}}$	$0.904_{\scriptstyle\pm.067}$	$0.713_{\scriptstyle\pm.025}$	$0.536_{\scriptstyle\pm.017}$	$0.420_{\scriptstyle\pm.018}$	$0.957_{\scriptstyle\pm.081}$	$0.768_{\scriptstyle\pm.044}$	$\underline{0.410_{\scriptstyle\pm.031}}$	$\underline{0.299_{\scriptstyle\pm.025}}$	$\mathbf{1.315_{\scriptstyle\pm.045}}$	$\mathbf{0.915_{\scriptstyle\pm.016}}$
	MERGE [ 9 ]	$0.609_{\scriptstyle\pm.023}$	$0.510_{\scriptstyle\pm.031}$	$1.082_{\scriptstyle\pm.247}$	$0.788_{\scriptstyle\pm.096}$	$0.483_{\scriptstyle\pm.051}$	$0.381_{\scriptstyle\pm.043}$	$0.998_{\scriptstyle\pm.160}$	$0.792_{\scriptstyle\pm.057}$	$0.242_{\scriptstyle\pm.016}$	$0.151_{\scriptstyle\pm.015}$	$1.531_{\scriptstyle\pm.023}$	$0.969_{\scriptstyle\pm.006}$

Table 1: Comparison on cSCC, Her2ST, and Kidney datasets using PCC-50, PCC-200, MSE, and MAE. Scores are averaged over test slices and three random seeds, reported as mean

\pm

standard deviation. Best results are in bold , and second-best are underlined .

Figure 2: Boxplots of per-slice PCC-50 (left) and PCC-200 (right) with a common random seed for (a) cSCC, (b) Her2ST, and (c) Kidney. Each point represents one test slice. Paired Wilcoxon signed-rank tests assess pairwise differences against baselines. * $p\text{-value}{<}0.05$ , ** $p\text{-value}{<}0.01$ , *** $p\text{-value}{<}0.001$ , ns not significant.

Figure 3: Expression of KRT6A on the P2_ST_rep3 slice from cSCC (a) and GNAS on the H1 slice from Her2ST (b). Each panel shows the ground truth and predictions from representative methods. Both genes are known markers with localized expression.

Figure 4: Gene–gene correlation matrices computed on the DKD Kidney slice 31-10042. Each matrix is derived from the predicted expression values of the HMHVG gene set. The ground truth and representative method outputs are shown.

Variant	PCC-50 $\uparrow$	PCC-200 $\uparrow$	MSE $\downarrow$	MAE $\downarrow$
Scratch	0.7425	0.6395	2.6175	1.1834
Decoder-Tune	0.8518	0.7699	1.3275	0.8999
Backbone-LoRA	0.8457	0.7618	1.2706	0.8800
HINGE	0.8755	0.8021	1.0096	0.7793

Table 2: Effect of reusing a pre-trained sc - FM under different adaptation schemes.

Variant	PCC-50 $\uparrow$	PCC-200 $\uparrow$	MSE $\downarrow$	MAE $\downarrow$
Gauss-Diff	0.7738	0.6415	1.7249	0.9855
Mask-Diff (NoCurr)	0.8691	0.7932	1.1619	0.8483
Mask-Diff (RandMask)	0.8752	0.7996	1.0208	0.7869
Mask-Diff (HINGE)	0.8755	0.8021	1.0096	0.7793

Table 3: Comparison of Gaussian diffusion, masked diffusion variants, and the full HINGE objective.

Variant	PCC-50 $\uparrow$	PCC-200 $\uparrow$	MSE $\downarrow$	MAE $\downarrow$
Hist-Affine-LN	0.8298	0.7690	1.7371	0.9809
SoftAdaLN (NoSoftNorm)	0.8658	0.7874	1.2317	0.8699
SoftAdaLN (NoIdInit)	0.8631	0.7722	1.2071	0.8600
SoftAdaLN (Full)	0.8755	0.8021	1.0096	0.7793

Table 4: Comparison of alternative conditioning mechanisms.

Variant	PCC-50 $\uparrow$	PCC-200 $\uparrow$	MSE $\downarrow$	MAE $\downarrow$
UNI	0.8625	0.7871	1.1512	0.8406
CONCH	0.8613	0.7783	1.3316	0.9029
UNI + CONCH	0.8755	0.8021	1.0096	0.7793

Table 5: Impact of different histology encoders.

Figure S1 : Selected genes used across different datasets.

Figure S2 : Spatial gene expression predictions for KRT6A , KRT10 , and GJB2 on the cSCC dataset.

Figure S3 : Spatial gene expression predictions for GNAS , ERBB2 , and FASN on the Her2ST dataset.

Figure S4 : Spatial gene expression predictions for FXYD2 , ATP1B1 , and PODXL on the Kidney dataset.

Figure S5 : Gene-wise SSIM with marker genes highlighted.

Figure S6 : Gene–gene correlation heatmaps on the Kidney dataset.

Figure S7 : Step-wise analysis of masked diffusion sampling.

Figure S8 : Masking horizon $T$ . Evaluation metrics as functions of the masking horizon $T$ on representative slices from the cSCC, Her2ST, and Kidney datasets.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Platform	Patient	Tissue	Condition	Samples	Inter-spot Dist ( $\mu$ m)	Spot Diameter ( $\mu$ m)	Spots under Tissue	Genes per slice	Preservation	Reference
Her2ST	ST	Patient A	Breast	Cancer	6	200	100	325–360	15,045–15,645	Fresh Frozen	PMID: 34650042
	ST	Patient B	Breast	Cancer	6	200	100	270–295	15,109–15,387	Fresh Frozen	same as above
	ST	Patient C	Breast	Cancer	6	200	100	176–187	15,557–15,842	Fresh Frozen	same as above
	ST	Patient D	Breast	Cancer	6	200	100	301–315	15,396–15,666	Fresh Frozen	same as above
	ST	Patient E	Breast	Cancer	3	200	100	570–587	15,097–15,701	Fresh Frozen	same as above
	ST	Patient F	Breast	Cancer	3	200	100	691–712	14,861–15,067	Fresh Frozen	same as above

Appendix

Table S1 : Summary of spatial transcriptomics datasets aggregated by patient. Each patient entry reports spot- and gene-level ranges across multiple slices, together with the corresponding platform.

Variant	PCC-50 $\uparrow$	PCC-200 $\uparrow$	MSE $\downarrow$	MAE $\downarrow$
Scratch	0.2511	0.1804	1.9176	1.1215
Decoder-Tune	0.4726	0.3374	0.9814	0.7716
Backbone-LoRA	0.4599	0.3163	0.9975	0.7774
HINGE	0.4801	0.3355	0.9481	0.7638
Gauss-Diff	0.3437	0.2084	1.8403	1.0447
Mask-Diff (NoCurr)	0.4702	0.3203	0.9641	0.7691

Appendix

Table S2 : Ablations on Her2ST. Component-wise analysis of HINGE variants on the Her2ST (A1) dataset.

Variant	PCC-50 $\uparrow$	PCC-200 $\uparrow$	MSE $\downarrow$	MAE $\downarrow$
Scratch	0.2517	0.2072	1.7576	1.1192
Decoder-Tune	0.4702	0.3627	0.9066	0.7552
Backbone-LoRA	0.4558	0.3271	1.0054	0.7984
HINGE	0.4871	0.3815	0.8956	0.7467
Gauss-Diff	0.2592	0.1673	1.7559	1.0054
Mask-Diff (NoCurr)	0.4725	0.3735	0.9086	0.7591

Appendix

Table S3 : Ablations on Kidney. Component-wise analysis of HINGE variants on the Kidney (IU-F52) dataset.

Methods	cSCC				Her2ST				Kidney
Methods	PCC-50 $\uparrow$	PCC-200 $\uparrow$	MSE $\downarrow$	MAE $\downarrow$	PCC-50 $\uparrow$	PCC-200 $\uparrow$	MSE $\downarrow$	MAE $\downarrow$	PCC-50 $\uparrow$	PCC-200 $\uparrow$	MSE $\downarrow$	MAE $\downarrow$
Linear ( $\zeta{=}1$ )	0.8728	0.7957	1.2067	0.8627	0.4922	0.3524	1.0802	0.8220	0.4817	0.3751	0.8986	0.7495
Cosine	0.8641	0.7868	1.1760	0.8392	0.4652	0.3294	1.1293	0.8383	0.4579	0.3469	1.0026	0.7999
$\zeta=0.5$	0.8743	0.7964	1.2196	0.8576	0.5090	0.3578	1.1948	0.8619	0.4567	0.3485	0.9886	0.7948
$\zeta=2$	0.8657	0.7828	1.3044	0.8990	0.4913	0.3450	1.0377	0.7888	0.4773	0.3729	0.9504	0.7829
$\zeta=\log_{T}G$	0.8755	0.8021	1.0096	0.7793	0.4817	0.3751	0.8976	0.7445	0.4871	0.3815	0.8956	0.7467

Appendix

Table S4 : Masking schedules. Results under different masking schedules on representative slices from the cSCC (P2_ST_rep3), Her2ST (A1), and Kidney (IU-F52) datasets.

Method	Type	Steps	Time/slide (s) $\downarrow$	spots/s $\uparrow$	PCC@50 $\uparrow$	Time@HR (s) $\downarrow$
ST-Net	Reg.	1	0.7162	890.82	0.739	13.97
BLEEP	Reg.	1	48.1776	13.24	0.785	991.29
TRIPLEX	Reg.	1	10.8633	58.45	0.805	OOM
Stem	Gen.	1000	230.1942	2.77	0.823	4138.51
STFlow	Gen.	10	0.2431	2624.83	0.692	OOM
HINGE (scGPT)	Gen.	5	1.1643	547.97	0.806	37.3242

Appendix

Table S5 : Inference efficiency (OOM: Out of Memory).

Gene Expression Prediction from Histopathology Images

Gene expression prediction from histopathology images links H&E tissue morphology to molecular profiles, enabling virtual spatial transcriptomics without sequencing every sample.

See all 44 papers

Jul 6, 2026

DriftST: One-Step Generative Inference of Spatial Transcriptomics from H&E Histology

Spatial Transcriptomics (ST) measures gene expression while preserving spatial context, but its high cost and low throughput leave public datasets small. Inferring expression directly from widely available Hematoxylin and Eosin (H&E) stained histology offers a cost-effective alternative. However, existing approaches face several limitations: regression methods over-smooth toward the conditional mean, while generative methods are faithful but require slow multi-step inference; most methods treat genes as independent and equally important, ignoring inter-gene dependencies and heterogeneous gene informativeness; and most are tailored to a single resolution, either spot-level or cell-level. To address these issues, we propose DriftST, a unified framework for inferring spatially resolved gene expression from H&E images. DriftST builds on a Cellular Drifting generative model that learns a direct drift from a histology-conditioned source to the expression distribution, retaining generative expressiveness while enabling efficient one-step generation. To capture gene structure, we introduce the STransformer, which combines a co-expression attention module for inter-gene dependencies with a gene residual gate for differential gene importance. Operating on a generic gene-panel representation, DriftST applies directly to both spot-level and cell-level data in one framework, and extensive experiments across diverse tissues and platforms show that it achieves state-of-the-art performance at both resolutions.

Yuhang Yang, Yonggan Bu, Shengyuan Zhou +2

University of Science and Technology of China, Hefei, China · Peking University Cancer Hospital, Beijing, China

Computational Pathology Spatial Transcriptomics One-Step Generative Modeling Spatial Gene Expression Prediction Gene Expression Prediction from Histopathology Images

Ask PDF

Sep 29, 2026

Towards Scalable Context-Aware Single-Cell Spatial Transcriptomics Prediction from Histology Images

Predicting gene expression from H&E-stained histology images offers a scalable alternative to costly spatial transcriptomics, yet most existing methods operate at the spot level, where signals from multiple cells are aggregated and critical cellular heterogeneity is obscured. Extending this paradigm to single-cell resolution is non-trivial. Naively applying pathology foundation models faces a scale mismatch: their patch-level representations mix multiple cells, whereas per-cell cropping or resizing distorts morphology and removes local context. Conversely, segmentation-based models without strong pretrained visual encoders often lack the morphological representation capacity needed for accurate molecular prediction and inherit errors from imperfect cell boundary masks. Here, we present CELLO, an efficient end-to-end framework that performs a single pathology foundation model forward pass per image and uses grid sampling to extract location-specific features for all cells simultaneously. We further introduce a distance-decay cross-attention module that refines each cell representation using spatially biased local morphological context. Using 52 public Xenium-H&E pairs from HEST-1k that span 12 organs and approximately 10 million cells, CELLO improves the average predictive accuracy over the evaluated baselines while reducing the mean whole-slide inference time compared to DeepSpot2Cell, a 14.0x speed-up on average that excludes upstream cell segmentation. Our work establishes a scalable foundation for single-cell gene expression prediction from H&E images.

Zijun Gao, Chunbin Gu, Jinxi Xiang +2

The Chinese University of Hong Kong · Stanford University School of Medicine

Spatial Transcriptomics Spatial Attention Spatial Gene Expression Prediction Gene Expression Prediction from Histopathology Images Whole-Slide Image Analysis

Ask PDF

Jun 1, 2026

GC-MoE: Genomics-Guided Cell-Type-Specific Mixture of Experts for Histology-Based Single-Cell Spatial Transcriptomics

Histology-based single-cell spatial transcriptomics (ST) estimation aims to predict gene expression for individual cells from histopathological images and cell locations, reducing the need for costly single-cell ST measurements. Unlike existing histology-to-ST methods that mainly predict spot-level profiles for local regions containing multiple cells, this task requires modeling cell-to-cell expression variability, which is strongly structured by cell type. We propose Genomics-Guided Cell-Type-Specific Mixture-of-Experts (GC-MoE), which estimates cell-type probabilities with a routing network and softly combines cell-type-specific experts for gene expression prediction. To further encode cell-type-dependent gene programs, we introduce the Cell-Type-Specific Co-Expression-Aware Predictor (CAP), together with a lightweight Cell-to-Cell Interaction Attention (C2CA) module for neighboring-cell context. Experiments and ablations on public single-cell ST datasets show consistent improvements over existing single-cell and adapted spot-level baselines.

Kaito Shiku, Ahtisham Fazeel Abbasi, Ryoma Bise +4

Kyushu University, Japan · German Research Center for Artificial Intelligence (DFKI GmbH), Germany · RPTU University Kaiserslautern-Landau, Germany +3

Computational Pathology Spatial Transcriptomics Spatial Gene Expression Prediction Mixture-of-Experts Models Gene Expression Prediction from Histopathology Images Single-Cell Transcriptomics

Ask PDF

Abstract

Gene Expression Prediction from Histopathology Images

Explore similar work

DriftST: One-Step Generative Inference of Spatial Transcriptomics from H&E Histology

Towards Scalable Context-Aware Single-Cell Spatial Transcriptomics Prediction from Histology Images

GC-MoE: Genomics-Guided Cell-Type-Specific Mixture of Experts for Histology-Based Single-Cell Spatial Transcriptomics