cs.LGSep 23, 2026

SpaFactor: Lightweight Spatial Context-Aware Gene Program Modeling for Histology-to-Transcriptomics Inference

Authors: Shiting Ruan, Xitong Ling, Qiming He, Ziyou Yan, Huaitian Yuan, Tian Guan, Ying Xiao, Xu Guan, +1 more

Organizations: Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · Fuzhou University, Fuzhou, Fujian, China · Beijing Tsinghua Changgung Hospital, School of Clinical Medicine, Tsinghua University, Beijing, China · Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Spatial transcriptomics (ST) profiles gene expression within tissue architecture, but its cost and experimental complexity limit routine use. Predicting spatial expression from routinely available hematoxylin and eosin (HE) images therefore offers a scalable alternative. However, conventional methods often fit high-dimensional gene outputs as independent targets, overlooking the biological coordination among genes while remaining vulnerable to high-dimensional noise and overfitting. Existing attempts to address this limitation often rely on computationally heavy graph networks or complex auxiliary supervision. We therefore introduce SpaFactor, a lightweight and efficient low-rank morphology-program-gene factorization framework. At the input, SpaFactor efficiently fuses the visual representation of the central spot with multiscale local and regional neighborhood context, yielding a histologic representation that captures cellular morphology and microenvironmental heterogeneity. For modeling, a residual MLP stably learns a nonlinear mapping from the tissue microenvironment to low-dimensional latent gene programs. These activities are decoded through shared gene loadings into coordinated multi-gene expression predictions. Across five public cohorts, SpaFactor achieves the best aggregate performance, with particularly clear improvements for spatially variable genes, and more faithfully recovers biologically organized spatial patterns. These results demonstrate that lightweight joint modeling of tissue context and gene programs can improve both predictive accuracy and biological fidelity.

Figures & tables

Figure 1: Overview of SpaFactor. (a) The central spot’s frozen GigaPath embedding is concatenated with mean embeddings from its local ( k=4k=4 ) and regional ( k=16k=16 ) neighborhoods, mapped by a four-block residual MLP to latent program activities, and decoded through shared gene loadings. (b) Detailed residual encoder and factor decoder. (c) Optional SpaFactor-Cal adds an inner-fold out-of-fold calibration loss during training and is removed for H&E-only inference.
DatasetMethodSpot PCC (%)Gene PCC (%)VR-PCC@50 (%)VR-PCC@100 (%)VR-PCC@200 (%)
Primary breastStrongest prior (HE2RNA)71.19 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 3.37\hss}\phantom{00.00}}33.86 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 4.10\hss}\phantom{00.00}}68.03 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.26\hss}\phantom{00.00}}64.12 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.45\hss}\phantom{00.00}}59.49 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.69\hss}\phantom{00.00}}
 GigaPath-MLP70.29 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 3.60\hss}\phantom{00.00}}33.10 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 4.48\hss}\phantom{00.00}}67.41 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.61\hss}\phantom{00.00}}63.44 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.76\hss}\phantom{00.00}}58.84 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 7.01\hss}\phantom{00.00}}
 SpaFactor-Cal71.73 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 3.28\hss}\phantom{00.00}}36.33 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 3.71\hss}\phantom{00.00}}71.69 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.17\hss}\phantom{00.00}}67.79 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.00\hss}\phantom{00.00}}63.20 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.23\hss}\phantom{00.00}}
 SpaFactor71.67 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 3.29\hss}\phantom{00.00}}36.36 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 4.17\hss}\phantom{00.00}}72.05 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.55\hss}\phantom{00.00}}68.09 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.42\hss}\phantom{00.00}}63.41 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.68\hss}\phantom{00.00}}
Visium breastStrongest prior (hist2RNA)70.07 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 2.38\hss}\phantom{00.00}}25.27 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 5.69\hss}\phantom{00.00}}40.98 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 8.51\hss}\phantom{00.00}}42.74 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 7.68\hss}\phantom{00.00}}44.55 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 8.99\hss}\phantom{00.00}}
 GigaPath-MLP68.54 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 2.10\hss}\phantom{00.00}}24.39 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 4.34\hss}\phantom{00.00}}38.62 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.80\hss}\phantom{00.00}}40.63 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.09\hss}\phantom{00.00}}42.54 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 7.14\hss}\phantom{00.00}}
Table 1: Five-cohort comparison. Values are fivefold means with standard deviations in lower-right subscripts. For each dataset, “strongest prior” denotes the published baseline with the best mean rank across the five metrics, with Gene PCC breaking ties. Bold marks the highest displayed mean; complete per-model results are reported separately.
Figure 2: Primary-cohort comparison with seven published methods. Panels show fivefold mean and standard deviation for Spot PCC, Gene PCC, and VR-PCC@50/100/200. Percentages show SpaFactor’s relative gain over HE2RNA, the Table 1 strongest prior for this cohort: (SpaFactor−HE2RNA)/HE2RNA×100%(\text{SpaFactor}-\text{HE2RNA})/\text{HE2RNA}\times 100\% .
VariantSpot PCCGene PCCVR-PCC@50VR-PCC@100VR-PCC@200
Full SpaFactor71.67 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 3.29\hss}\phantom{00.00}}36.36 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 4.17\hss}\phantom{00.00}}72.05 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.55\hss}\phantom{00.00}}68.09 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.42\hss}\phantom{00.00}}63.41 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.68\hss}\phantom{00.00}}
Δ\Delta Spot only70.59 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 3.51\hss}\phantom{00.00}}33.99 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 3.96\hss}\phantom{00.00}}68.07 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.04\hss}\phantom{00.00}}64.10 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.22\hss}\phantom{00.00}}59.61 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.32\hss}\phantom{00.00}}
Δ\Delta Local only71.51 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 3.46\hss}\phantom{00.00}}35.66 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 4.11\hss}\phantom{00.00}}70.98 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.56\hss}\phantom{00.00}}67.00 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.43\hss}\phantom{00.00}}62.28 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.72\hss}\phantom{00.00}}
Δ\Delta Regional only71.66 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 3.36\hss}\phantom{00.00}}36.14 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 4.05\hss}\phantom{00.00}}71.88 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.25\hss}\phantom{00.00}}67.87 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.35\hss}\phantom{00.00}}63.12 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.59\hss}\phantom{00.00}}
Δ\Delta Plain MLP64.30 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 1.52\hss}\phantom{00.00}}16.74 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 13.59\hss}\phantom{00.00}}34.57 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 24.55\hss}\phantom{00.00}}32.86 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 23.69\hss}\phantom{00.00}}31.49 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 23.15\hss}\phantom{00.00}}
Δ\Delta Direct decoder71.50 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 3.30\hss}\phantom{00.00}}35.74 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 4.25\hss}\phantom{00.00}}71.61 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 7.10\hss}\phantom{00.00}}67.66 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.92\hss}\phantom{00.00}}62.91 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 7.09\hss}\phantom{00.00}}
Table 2: Structural component ablations on the primary benchmark. All metrics are percentages; values are fivefold means with standard deviations in lower-right subscripts.
KKSpot PCCGene PCCVR-PCC@50VR-PCC@100VR-PCC@200
6471.60 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 3.56\hss}\phantom{00.00}}36.18 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 4.15\hss}\phantom{00.00}}71.41 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.25\hss}\phantom{00.00}}67.45 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.51\hss}\phantom{00.00}}62.84 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.70\hss}\phantom{00.00}}
12871.67 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 3.50\hss}\phantom{00.00}}35.82 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 4.34\hss}\phantom{00.00}}71.40 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.54\hss}\phantom{00.00}}67.31 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.53\hss}\phantom{00.00}}62.53 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.80\hss}\phantom{00.00}}
25671.67 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 3.29\hss}\phantom{00.00}}36.36 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 4.17\hss}\phantom{00.00}}72.05 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.55\hss}\phantom{00.00}}68.09 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.42\hss}\phantom{00.00}}63.41 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.68\hss}\phantom{00.00}}
51271.55 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 3.32\hss}\phantom{00.00}}36.46 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 4.07\hss}\phantom{00.00}}72.38 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.14\hss}\phantom{00.00}}68.41 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.26\hss}\phantom{00.00}}63.72 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.56\hss}\phantom{00.00}}
102471.57 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 3.31\hss}\phantom{00.00}}36.06 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 4.29\hss}\phantom{00.00}}72.02 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.30\hss}\phantom{00.00}}67.99 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.46\hss}\phantom{00.00}}63.24 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.87\hss}\phantom{00.00}}
Table 3: Sensitivity to latent factor count on the primary benchmark. All metrics are percentages; values are fivefold means with standard deviations in lower-right subscripts.
Figure 3: Residual-depth sensitivity on the primary benchmark. Curves show fivefold mean Spot PCC (diamonds) and VR-PCC@50/100/200 (circles/squares/triangles); all metrics are percentages. Values are annotated at L=4L=4 , 10, and 12.
DepthTrainable parametersTraining time (s)Full-fold inference (s)
L=4L=422,304,97658.330.0315
L=10L=1047,501,52067.020.0365
L=12L=1255,900,36880.420.0398
Table 4: Controlled depth–efficiency comparison on one held-out primary-cohort fold.
λPCC\lambda_{\mathrm{PCC}}Spot PCCGene PCCVR-PCC@50VR-PCC@100VR-PCC@200
0.0071.59 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 3.32\hss}\phantom{00.00}}36.22 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 4.19\hss}\phantom{00.00}}71.95 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.56\hss}\phantom{00.00}}67.93 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.41\hss}\phantom{00.00}}63.27 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.61\hss}\phantom{00.00}}
0.0571.67 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 3.37\hss}\phantom{00.00}}35.99 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 4.08\hss}\phantom{00.00}}71.70 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.39\hss}\phantom{00.00}}67.69 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.34\hss}\phantom{00.00}}62.97 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.60\hss}\phantom{00.00}}
0.1071.67 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 3.29\hss}\phantom{00.00}}36.36 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 4.17\hss}\phantom{00.00}}72.05 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.55\hss}\phantom{00.00}}68.09 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.42\hss}\phantom{00.00}}63.41 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.68\hss}\phantom{00.00}}
0.2071.57 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 3.33\hss}\phantom{00.00}}36.23 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 4.10\hss}\phantom{00.00}}71.81 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.90\hss}\phantom{00.00}}67.86 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.64\hss}\phantom{00.00}}63.18 {}_{\scriptscriptstyle\pm\hbox to0.0pt{\scriptscriptstyle 6.81\hss}\phantom{00.00}}
Table 5: Gene-PCC loss-weight sensitivity on the primary benchmark. All metrics are percentages; values are fivefold means with standard deviations in lower-right subscripts. λPCC=0.10\lambda_{\mathrm{PCC}}=0.10 is the main setting.
Figure 4: Spatial recovery of four breast biomarkers on one held-out section. Rows show COL1A1, GATA3, ERBB2, and ESR1; columns show measured expression and predictions from SpaFactor (“Ours”), ST-Net, DeepSpot, and matched GigaPath-MLP on the same spots. Pearson rr measures agreement with the measured map, and each row shares a common color scale.
Figure 5: Spatial recovery of KEGG oxidative-phosphorylation activity on a held-out section. The H&E image, measured activity, and predictions from SpaFactor and matched GigaPath-MLP are aligned on the same 301 spots and share a zz -score color scale.
Figure 6: Pathway-activity fit on the held-out section in Figure 5 . Each point is one spot; solid lines are fitted regressions and dashed lines indicate identity. SpaFactor is shown on the left and matched GigaPath-MLP on the right; PCC, MAE, and RMSE are reported in each panel.
Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

CohortSectionsSpotsPlatform
Primary breast10845,306Spatial Transcriptomics
Visium breast87163,931Visium
Bowel73205,642Visium
Brain91322,572Visium
Skin6261,634Visium
Total421799,085N/A
Appendix
Table A.1: Cohorts used in the anonymous submission.
ComponentSetting
Histology encoderFrozen Prov-GigaPath, 1,536 dimensions
ContextCentral spot + mean embeddings within the slide for k=4k=4 and k=16k=16
PredictorResidual MLP, hidden width 1,024, inner width 2,048, 4 blocks
Decoder256 factors with a shared gene loading matrix
RegularizationDropout 0.10; AdamW weight decay 10−510^{-5}
OptimizationAdamW; learning rate 10−310^{-3} ; batch size 1,024; gradient clip 5.0; at most 100 epochs
Appendix
Table A.2: Final SpaFactor configuration without a teacher, used for the main results.
MetricMean delta (pp)Wins95% CI (pp)
Spot PCC+1.0223/25[0.63, 1.45]
Gene PCC+2.0920/25[0.80, 3.44]
VR-PCC@50+3.3721/25[2.07, 4.67]
VR-PCC@100+3.0621/25[1.62, 4.44]
VR-PCC@200+2.8221/25[1.27, 4.25]
Appendix
Table A.3: Final SpaFactor minus matched GigaPath-MLP across five cohorts and five folds.
SeedSpot PCC (%)Gene PCC (%)VR-PCC@50 (%)VR-PCC@100 (%)VR-PCC@200 (%)
171.52 ±\pm 3.3736.51 ±\pm 3.9372.08 ±\pm 6.1268.24 ±\pm 6.0463.60 ±\pm 6.35
1071.62 ±\pm 3.2735.94 ±\pm 3.7571.76 ±\pm 5.7867.72 ±\pm 5.8863.01 ±\pm 5.99
4271.67 ±\pm 3.2936.36 ±\pm 4.1772.05 ±\pm 6.5568.09 ±\pm 6.4263.41 ±\pm 6.68
123471.65 ±\pm 3.3036.06 ±\pm 3.9471.61 ±\pm 6.6767.60 ±\pm 6.6162.86 ±\pm 6.76
340771.66 ±\pm 3.3635.93 ±\pm 4.2771.35 ±\pm 6.7567.33 ±\pm 6.8062.60 ±\pm 7.01
Appendix
Table C.1: Initialization stability across five training seeds; PCC mean ±\pm standard deviation over five folds (%).
MethodSpot PCC (%)Gene PCC (%)VR-PCC@50 (%)VR-PCC@100 (%)VR-PCC@200 (%)
ST-Net70.45 ±\pm 3.8335.10 ±\pm 6.8967.29 ±\pm 8.5662.43 ±\pm 8.6956.10 ±\pm 8.99
BLEEP65.09 ±\pm 3.1018.88 ±\pm 3.7847.97 ±\pm 8.1843.38 ±\pm 8.9437.34 ±\pm 8.14
DeepSpot69.11 ±\pm 3.2927.92 ±\pm 2.9065.70 ±\pm 6.4459.63 ±\pm 7.2151.65 ±\pm 7.15
GenST64.74 ±\pm 3.5216.74 ±\pm 9.6440.35 ±\pm 24.5236.42 ±\pm 22.0931.40 ±\pm 19.13
HisToGene68.55 ±\pm 3.3525.72 ±\pm 3.0764.79 ±\pm 7.0258.66 ±\pm 7.6950.77 ±\pm 7.56
HE2RNA71.19 ±\pm 3.3733.86 ±\pm 4.1068.03 ±\pm 6.2664.12 ±\pm 6.4559.49 ±\pm 6.69
Appendix
Table D.1: Complete source normalized benchmark for the primary breast cohort. Values are fivefold mean ±\pm standard deviation (%); bold marks the highest mean in each column.
MethodSpot PCC (%)Gene PCC (%)VR-PCC@50 (%)VR-PCC@100 (%)VR-PCC@200 (%)
ST-Net69.85 ±\pm 3.0348.40 ±\pm 6.8438.91 ±\pm 7.6839.28 ±\pm 7.8639.95 ±\pm 8.84
BLEEP68.37 ±\pm 2.5013.58 ±\pm 2.9626.61 ±\pm 5.4126.10 ±\pm 3.9724.13 ±\pm 5.20
DeepSpot70.46 ±\pm 2.3621.58 ±\pm 4.2537.21 ±\pm 4.0635.95 ±\pm 4.5334.23 ±\pm 5.70
GenST70.21 ±\pm 2.4619.43 ±\pm 3.0231.82 ±\pm 4.5831.21 ±\pm 4.7830.20 ±\pm 5.09
HisToGene69.07 ±\pm 2.4218.93 ±\pm 4.1737.41 ±\pm 3.8435.43 ±\pm 4.8533.86 ±\pm 6.32
HE2RNA69.88 ±\pm 2.4824.85 ±\pm 4.5639.56 ±\pm 7.0141.78 ±\pm 6.0043.89 ±\pm 7.19
Appendix
Table D.2: Complete benchmark for the Visium breast cohort. Values are fivefold mean ±\pm standard deviation (%); bold marks the highest mean in each column.
MethodSpot PCC (%)Gene PCC (%)VR-PCC@50 (%)VR-PCC@100 (%)VR-PCC@200 (%)
ST-Net53.59 ±\pm 6.9836.87 ±\pm 21.6847.88 ±\pm 25.9246.42 ±\pm 24.6443.38 ±\pm 23.77
BLEEP50.71 ±\pm 5.3715.76 ±\pm 6.6430.87 ±\pm 12.3227.67 ±\pm 11.1623.82 ±\pm 9.88
DeepSpot57.20 ±\pm 5.6731.15 ±\pm 6.5650.97 ±\pm 8.9746.73 ±\pm 8.3442.26 ±\pm 7.78
GenST55.25 ±\pm 5.5227.02 ±\pm 6.2644.35 ±\pm 8.8140.65 ±\pm 8.3536.62 ±\pm 7.81
HisToGene54.96 ±\pm 5.5025.17 ±\pm 7.0045.25 ±\pm 10.1241.05 ±\pm 9.7637.00 ±\pm 8.89
HE2RNA57.64 ±\pm 5.2043.78 ±\pm 6.2660.20 ±\pm 7.9758.46 ±\pm 7.3355.76 ±\pm 6.91
Appendix
Table D.3: Complete bowel benchmark. Values are fivefold mean ±\pm standard deviation (%); bold marks the highest mean in each column.
MethodSpot PCC (%)Gene PCC (%)VR-PCC@50 (%)VR-PCC@100 (%)VR-PCC@200 (%)
ST-Net60.07 ±\pm 2.4027.66 ±\pm 26.1940.18 ±\pm 20.1936.71 ±\pm 19.5232.23 ±\pm 18.82
BLEEP55.73 ±\pm 1.5916.28 ±\pm 3.6834.10 ±\pm 11.7729.89 ±\pm 9.4824.31 ±\pm 8.48
DeepSpot57.94 ±\pm 1.7922.07 ±\pm 2.2941.74 ±\pm 8.7836.96 ±\pm 6.6331.18 ±\pm 5.51
GenST57.58 ±\pm 2.2220.84 ±\pm 3.0239.48 ±\pm 9.6535.19 ±\pm 7.6329.56 ±\pm 6.90
HisToGene56.93 ±\pm 2.4218.47 ±\pm 3.6238.31 ±\pm 11.8533.95 ±\pm 9.2428.33 ±\pm 8.02
HE2RNA60.84 ±\pm 2.4530.16 ±\pm 6.0851.73 ±\pm 8.7648.30 ±\pm 8.3943.41 ±\pm 8.15
Appendix
Table D.4: Complete brain benchmark. Values are fivefold mean ±\pm standard deviation (%); bold marks the highest mean in each column.
MethodSpot PCC (%)Gene PCC (%)VR-PCC@50 (%)VR-PCC@100 (%)VR-PCC@200 (%)
ST-Net66.36 ±\pm 6.7663.04 ±\pm 6.1674.32 ±\pm 9.7274.22 ±\pm 10.0171.55 ±\pm 9.96
BLEEP62.70 ±\pm 6.4413.73 ±\pm 5.5439.23 ±\pm 12.0236.41 ±\pm 8.8433.32 ±\pm 9.16
DeepSpot65.32 ±\pm 6.2824.76 ±\pm 1.9652.59 ±\pm 7.4449.92 ±\pm 6.3547.98 ±\pm 4.63
GenST64.49 ±\pm 5.7421.36 ±\pm 2.9146.45 ±\pm 10.1345.11 ±\pm 7.9543.33 ±\pm 6.32
HisToGene64.93 ±\pm 6.0320.73 ±\pm 1.9652.59 ±\pm 8.1349.76 ±\pm 7.0147.45 ±\pm 5.32
HE2RNA68.05 ±\pm 6.3151.08 ±\pm 4.7879.19 ±\pm 6.0979.21 ±\pm 6.2878.13 ±\pm 5.99
Appendix
Table D.5: Complete skin benchmark. Values are fivefold mean ±\pm standard deviation (%); bold marks the highest mean in each column.

Explore similar work

CardsList
Jul 27, 2026cs.CV

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

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.

Ziang Liu, Xinhai Chen, Yigui Feng +3

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 TranscriptomicsDigital PathologyTissue
AskPDF
Jul 10, 2026cs.LG

COAST: Context-Aware Differential Learning for Gene Expression Prediction in Spatial Transcriptomics

Spatial transcriptomics enables profiling of spatial gene expression but is limited by high cost and low throughput, motivating prediction from H&E histopathology images. Existing context-aware methods mainly supervise absolute expression, while relative expression relationships between spots are rarely used explicitly. We propose COAST, a context-aware differential learning framework for spatial gene expression prediction. COAST conditions the local and global context features with type-specific modulation and aggregates the target and context spot tokens using a Transformer encoder to capture both fine-grained local patterns and slide-level structure. It is trained with a joint objective that combines absolute expression regression with signed differential regression between the target and context spots. Experiments on multiple spatial transcriptomics datasets show consistent improvements in correlation- and distribution-based metrics, demonstrating the effectiveness of context-aware differential learning for histology-based spatial gene expression prediction.

Keunho Byeon, Sunhong Park, Jeewoo Lim +1

School of Electrical Engineering, Korea University, Seoul 02841, Republic of Korea

Spatial TranscriptomicsTranscriptomicsDigital PathologyOcean
AskPDF
Jul 6, 2026cs.CV

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

Spatial TranscriptomicsDigital PathologySpatio-Temporal TransformersGenerative Models
AskPDF