FEAST: Fully Connected Expressive Attention for Spatial Transcriptomics

Taejin Jeong; Joohyeok Kim; Jinyeong Kim; Chanyoung Kim; Seong Jae Hwang

cs.CVMar 26, 2026

FEAST: Fully Connected Expressive Attention for Spatial Transcriptomics

Authors: Taejin Jeong, Joohyeok Kim, Jinyeong Kim, Chanyoung Kim, Seong Jae Hwang

Organizations: Yonsei University · Emory University

Abstract

Spatial Transcriptomics (ST) provides spatially-resolved gene expression, offering crucial insights into tissue architecture and complex diseases. However, its prohibitive cost limits widespread adoption, leading to significant attention on inferring spatial gene expression from readily available whole slide images. While graph neural networks have been proposed to model interactions between tissue regions, their reliance on pre-defined sparse graphs prevents them from considering potentially interacting spot pairs, resulting in a structural limitation in capturing complex biological relationships. To address this, we propose FEAST (Fully connected Expressive Attention for Spatial Transcriptomics), an attention-based framework that models the tissue as a fully connected graph, enabling the consideration of all pairwise interactions. To better reflect biological interactions, we introduce negative-aware attention, which models both excitatory and inhibitory interactions, capturing essential negative relationships that standard attention often overlooks. Furthermore, to mitigate the information loss from truncated or ignored context in standard spot image extraction, we introduce an off-grid sampling strategy that gathers additional images from intermediate regions, allowing the model to capture a richer morphological context. Experiments on public ST datasets show that FEAST surpasses state-of-the-art methods in gene expression prediction while providing biologically plausible attention maps that clarify positive and negative interactions. Our code is available at https://github.com/starforTJ/ FEAST.

Figures & tables

Figure 1: Conceptual overview of our proposed framework, illustrating how it addresses key challenges of prior ST methods. Existing approaches are limited by sparse graphs that miss potential interactions, standard positive-only relationships in attention that omit inhibitory interactions, and sparse information that leads to information loss in off-grid regions. Our framework overcomes these limitations by employing a fully connected graph to consider all-pairs interactions, introducing a novel negative-aware attention to explicitly model both positive and inhibitory relationships, and utilizing pseudo-spots to capture and leverage off-grid information.

Figure 2: The overall architecture of FEAST. The framework first extracts features and spot distance ( $\mathbf{B}_{h}$ ) from the input WSI. The features are then processed through $L$ stacked hierarchical attention layers. Each layer consists of two stages: (1) a FEAST Block applied to local $k$ -nearest neighbors (including pseudo-spots) to absorb spatial context, followed by (2) a second FEAST Block applied to the original spots for global self-attention. The final representations are passed to an MLP head for gene expression prediction.

Figure 3: Detailed architecture of the FEAST Block. This block, which serves as the core attention mechanism in our framework, explicitly models both positive (excitatory) and negative (inhibitory) relationships. The raw feature score ( $qk^{T}$ ) and its inversion ( $-qk^{T}$ ) are used to compute positive and negative scores, respectively, both incorporating the same positional bias ( $\mathbf{B}_{h}$ ). After softmax normalization ( $\sigma$ ), the final attention weight is computed by subtracting the negative weight from the positive weight. This results in an attention map that can capture both positive (red) and negative (blue) biological relationships.

Figure 4: Limitations of the standard fixed-size patch extraction and the motivation for our off-grid sampling. (a) Patch Gaps (Missed Context): Sparse sampling creates off-grid gaps between original patches, causing critical context to be ignored. (b) Truncated Morphological Structures: A single, coherent morphological structure is arbitrarily truncated by the grid boundaries. Our proposed pseudo-spots are sampled from these information-loss regions to restore the complete context.

Method	ST-Net			Her2ST			SCC
Method	MSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$	MSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$	MSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$
ResNet+FCN	0.1999	0.3448	0.5221	0.6623	0.6385	0.4629	0.6103	0.6290	0.4619
BLEEP [ 27 ]	0.3756	0.4736	0.0784	0.7426	0.6591	0.2747	0.6079	0.6013	0.4176
HisToGene [ 19 ]	0.3054	0.4336	0.1211	0.9452	0.7739	0.2062	0.3095	0.4367	0.1225
Hist2ST [ 29 ]	0.3811	0.4822	0.1525	0.7843	0.7286	0.2479	1.0190	0.7639	0.3003
THItoGene [ 14 ]	0.2925	0.4111	0.3666	0.8436	0.7069	0.3445	0.6798	0.6442	0.3897

Table 1: Cross-validation performance on the ST-Net [ 10 ] , Her2ST [ 1 ] and SCC [ 13 ] datasets including baselines. The best results are in bold , and the second best are underlined . The performance for all baseline models are cited directly from the MERGE [ 9 ] .

Figure 5: Qualitative comparison of predicted gene expression heatmaps for FEAST and MERGE. Each row presents a side-by-side comparison for the cancer-relevant genes FASN and GNAS, showing the WSI, the ground truth gene expression, and the model predictions. The PCC scores in parentheses confirm that our model attains a higher correlation with the ground truth expressions.

Figure 6: Illustration of the effect of pseudo-spots (off-grid sampling) on target spot prediction. The PCC value in parentheses represents the prediction accuracy for this target spot (green box). The original spots surrounding the target spot (red boxes) are spatially sparse, failing to adequately capture the target’s information (w/o pseudo-spot). The PCC value improves when pseudo-spots are incorporated (Ours). This result emphasizes the necessity of our approach when surrounding information is sparse.

Figure 7: Comparison of attention maps with and without negative-aware attention. Basic attention (bottom) exhibited limitations, either misidentifying structurally different spots as positive or failing to capture them entirely. In contrast, when negative-aware attention was applied (top), it accurately identified structurally dissimilar spots as negative. Notably, we can observe that it clearly distinguishes structural differences even for spatially close spots, where basic attention had failed.

$k$	MSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$
8	0.5821	0.5819	0.5453
16	0.5796	0.5784	0.5504
32	0.5761	0.5782	0.5524
64	0.5793	0.5818	0.5481
100	0.5768	0.5807	0.5484

Table 2: Performance variation according to the value of

k

. The model achieved the best performance at

k=32

Negative	Off-grid	MSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$
		0.5878	0.5875	0.5396
✓		0.5778	0.5829	0.5464
	✓	0.5829	0.5831	0.5458
✓	✓	0.5761	0.5782	0.5524

Table 3: Ablation study on negative-aware attention and off-grid sampling strategy. The results confirm that applying these methods yields a performance gain compared to when they are not used.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Architecture
Config	Value
Input dimension ( $d$ )	1536
Number of attention heads (H)	8
Number of hierarchical layers ( $L$ )	3
Local neighborhood size ( $k$ )	32
Negative-aware Attention

Appendix

Table 4: Hyperparameter configuration for FEAST training.

Dataset	250 genes (Total genes to be predicted)
ST-Net	RPS3, IGLL5, RPLP1, TFF3, RPS18, GAPDH, TMSB10, RPLP2, RPS14, RPL37A, RPS19, RPL28, KRT19, RPL8, RPL13, RPL19, ACTB, RPL36, RPL18A, RPL35, RPL18, RPS2, RPS12, RPS21, RACK1, RPL13A, CTSD, FTL, PFN1, MGP, RPS15, RPS11, RPS16, HLA-B, UBA52, NHERF1, RPS17, PSAP, RPLP0, SERF2, RPS27, RPS8, RPL27A, MUC1, RPS28, H2AJ, RPL10, CALR, RPS29, RPL38, RPL11, P4HB, RPS6, CST3, FTH1, RPS4X, SSR4, RPL30, ERBB2, APOE, AZGP1, RPL3, COX6C, HLA-C, FAU, RPS9, EEF2, B2M, RPS5, RPL12, ACTG1, RPS27A, RPL37, RPL23, HLA-A, RPL31, RPL29, RPL7A, IFI27, PABPC1, CD74, BEST1, RPL32, FASN, S100A9, GPX4, RPL15, RPL27, MZT2B, RPL23A, HSPB1, MALAT1, RPS24, COL1A1, C4B, KRT18, CFL1, CD81, ALDOA, RPL35A, SYNGR2, PPP1CA, HLA-E, TAGLN, RPL9, CD63, RPS3A, LGALS3BP, IGFBP2, BST2, TPT1, EDF1, RPS25, ATP6V0B, TAPBP, GRINA, XBP1, S100A11, NBEAL1, AEBP1, CCND1, OAZ1, RPL14, TAGLN2, FN1, PPDPF, BCAP31, IFITM3, PRDX1, BGN, GNAS, PTMA, UBC, MZT2A, SLC25A6, RPS20, HSP90AB1, RPS10, MYL6, CLDN3, ATP6AP1, PRDX2, RPL24, GNB2, RPL34, RPL4, LMNA, NDUFA13, HLA-DRA, SNHG25, TIMP1, H1-0, RPS23, COX8A, KRT8, LY6E, ENO1, GRN, PTPRF, RPL7, UBB, BSG, ELOB, COX6B1, TMSB4X, C1QA, PRSS8, RPL5, UQCR11, RPS7, A2M, RPS15A, VIM, S100A6, NDUFA11, PSMD3, EVL, APOC1, H3-3B, ATP5F1E, PLXNB2, MYL9, TUBA1B, CTSB, ISG15, FLNA, RPS13, NDUFB9, EIF4A1, POLR2L, CYBA, CRIP2, EEF1D, ATP1A1, ELF3, TUFM, SH3BGRL3, STARD10, C3, GUK1, ZNF90, C12orf57, TLE5, SEC61A1, SDC1, PLD3, SPDEF, ARHGDIA, IFI6, LAPTM5, RPL41, CLU, GNAI2, PFDN5, RPL39, SSR2, COX4I1, RHOC, JUP, EIF4G1, FXYD3, TSPO, UQCRQ, COL1A2, RPL10A, S100A8, SELENOW, TPI1, ATP5MC2, PTMS, IGFBP5, LGALS1, SPINT2, RPSA, GSTP1, CHCHD2, EIF5A, COX5B, ATG10, RPL6, EEF1A1, CAPNS1, LMAN2, UBE2M, SPARC, EIF3C, GAS5, TUBB, ACTN4, IGFBP4
Her2ST	IGKC, TMSB10, ERBB2, IGHG3, IGLC2, IGHA1, GAPDH, ACTB, IGLC3, IGHM, SERF2, PSMB3, PFN1, ACTG1, KRT19, RACK1, MUCL1, CISD3, APOE, MIEN1, SSR4, CALR, PSAP, CTSD, FTL, FTH1, TPT1, PTPRF, UBA52, P4HB, BEST1, HLA-B, FAU, SLC9A3R1, FN1, COL1A1, EEF2, IGHG4, CALML5, CD74, B2M, FASN, S100A9, MGP, CFL1, PSMD3, IGHG1, HLA-A, S100A6, MYL6, COL1A2, PHB, TAGLN2, HLA-E, HLA-C, KRT7, CD63, SYNGR2, STARD3, PABPC1, GPX4, GRB7, SLC25A6, AEBP1, GNAS, NDUFB9, EDF1, CRIP2, DDX5, OAZ1, EIF4G1, LMNA, GNB2, CST3, PCGF2, SDC1, S100A11, PRDX1, GRINA, ATP6V0B, TFF3, HLADRA, EEF1D, AZGP1, PPP1CA, FLNA, COL3A1, ATP5E, SPDEF, AP000769.1, ALDOA, PLXNB2, TAGLN, TUBA1B, APOC1, PRRC2A, LAPTM5, PTMS, KRT18, IFI27, PLD3, ADAM15, C1QA, AES, TSPO, MLLT6, TAPBP, SCAND1, ATP1A1, CD81, SEC61A1, CLDN3, PPDPF, S100A14, BGN, C3, MZT2B, S100A8, MDK, PFDN5, H2AFJ, SH3BGRL3, ENO1, XBP1, CYBA, COX6B1, TRAF4, CD24, PRSS8, MMP14, MUC1, VIM, MIDN, SPINT2, BST2, TIMP1, GUK1, ACTN4, CTSB, COX4I1, CCT3, HNRNPA2B1, SEPW1, LY6E, SCD, HSPB1, EIF4G2, BSG, ZYX, TUBB, LASP1, CD99, COL6A2, H1FX, RALY, UBE2M, SPARC, ATG10, HSP90AB1, ORMDL3, LMAN2, CHCHD2, COX7C, ARHGDIA, VMP1, UBC, IGFBP2, COPE, NUPR1, PERP, KRT81, PPP1R1B, LGALS3BP, SSR2, KIAA0100, MYL9, CIB1, IDH2, STARD10, LGALS1, COX6C, GRN, MAPKAPK2, GNAI2, KDELR1, COL18A1, UQCRQ, COX5B, ELOVL1, CHPF, CLDN4, C12orf57, LGALS3, HSP90AA1, JUP, A2M, NDUFB7, PGAP3, HSPA8, TCEB2, PEBP1, COPS9, ATP5G2, ATP6AP1, MYH9, LSM4, COX8A, UQCR11, ATP5B, DHCR24, PTBP1, EIF3B, NDUFA3, FKBP2, MMACHC, RABAC1, ISG15, PTMA, RRBP1, POSTN, C1QB, BCAP31, PSMB4, LAPTM4A, INTS1, FNBP1L, JTB, NBL1, HM13, SLC2A4RG, ROMO1, SERINC2, NDUFA11, RHOC, TXNIP, TYMP, NACA, HSP90B1, SNRPB, PFKL, VCP, ERGIC1, NUCKS1, PSMD8, CALM2, AP2S1, DBI, C4orf48, SDF4, TPI1
SCC	S100A8, KRT6A, KRT14, S100A9, KRT5, KRT6B, KRT16, KRT6C, KRT17, MT-CO3, S100A7, MT-CO2, SFN, S100A2, MT-CO1, ACTB, PERP, SPRR1B, KRT10, KRT1, EEF1A1, RPLP1, LGALS7B, LGALS7, COL1A1, FABP5, RPS12, HLA-B, MT-ND4, RPLP2, ACTG1, GJB2, B2M, TPT1, RPL13, MT-ATP6, RPS24, PFN1, KRTDAP, RPS6, DMKN, RPLP0, MT-ND3, RPL37A, DSP, CXCL14, RPS18, RPS17, RPS8, RPL13A, MT-CYB, RPL11, RPL27A, RPL28, MT-ND1, RPS27, RPL32, CSTA, RPL34, RPL31, COL1A2, RPL8, SBSN, TMSB10, ENO1, RPS14, RPL36, SPRR2A, RPL39, GSTP1, RPS27A, JUP, RPS19, RPL37, RPL27, RPL3, RPS29, COL3A1, RPS11, CSTB, RPL9, RACK1, ANXA2, RPL7A, RPL23, RPL19, S100A11, RPS2, RPS28, EEF2, ANXA1, CD74, PABPC1, LDHA, RPS3, RPL35A, DSC2, AQP3, RPS25, IFI27, CALML5, YWHAZ, RPL6, TMSB4X, RPS23, RPL12, S100A14, RPS4X, UBA52, SLPI, PKP1, RPL38, HLA-A, RPS13, LY6D, RPL24, ATP1B3, MYL6, GJB6, S100A6, HSPB1, RPL18, MT-ND2, SDC1, IVL, FTL, RPS3A, RPL10, RPS15A, PI3, RPL18A, S100A10, RPS7, S100A7A, RPL29, RPL26, RPL41, RPL4, RPL7, SPARC, VIM, PTMA, RPS20, MMP1, SH3BGRL3, RPL15, MYH9, GJA1, ITM2B, PPIA, RPL14, UBC, RPL5, CD44, AHNAK, RPL21, DSC3, CNFN, CD24, CFL1, COL17A1, HSP90AA1, RPS16, PKM, NACA, RPS5, ALDOA, H3F3B, S100A16, TAGLN2, HLA-C, TRIM29, LYPD3, FAU, LMNA, SPINK5, SPRR2E, RPL22, KRT2, CST3, DSG3, CLCA2, RPSA, DSG1, RPS9, NDRG1, AC090498.1, GRN, TXN, HSPA8, TGFBI, CTSB, SPRR2D, HLA-DRA, ACTN4, RPS21, EIF1, CTSD, ARPC2, CALML3, KLK7, CALM1, GNAS, DYNLL1, FLG, FLNA, DST, SLC2A1, PSAP, EIF4G2, EEF1B2, FGFBP1, LGALS1, ITGA6, MYL12B, TPI1, RPL10A, TMEM45A, BTF3, DSTN, RTN4, HNRNPA2B1, LAD1, ATP1A1, SERPINB3, PRDX1, COL6A1, ATP5E, PPDPF, TYMP, CD63, EIF5A, YWHAQ, PGK1, HLA-E, IFITM3, RPS26, IGFBP4, OAZ1, NPM1, LCE3D, FXYD3, MT2A, COL6A2, POLR2L, CD59, HNRNPK, RPL35, TMBIM6, HSP90AB1

Appendix

Table 5: List of 250 target genes selected for each dataset.

Negative	Off-grid	ST-Net			Her2ST			SCC
Negative	Off-grid	MSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$	MSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$	MSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$
		0.1156	0.2700	0.7200	0.6538	0.6505	0.4920	0.5010	0.5730	0.5908
✓		0.1144	0.2696	0.7267	0.6321	0.6401	0.4958	0.4955	0.5522	0.5978
	✓	0.1126	0.2607	0.7222	0.6308	0.6397	0.5020	0.5035	0.5575	0.5850
✓	✓	0.1111	0.2603	0.7224	0.6158	0.6342	0.5194	0.4923	0.5462	0.5971

Appendix

Table 6: Ablation study on negative-aware attention and off-grid sampling strategy. The best results are in bold .

$k$	ST-Net			Her2ST			SCC
$k$	MSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$	MSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$	MSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$
8	0.1138	0.2630	0.7176	0.6327	0.6440	0.4957	0.4920	0.5438	0.5975
16	0.1119	0.2597	0.7237	0.6211	0.6366	0.5118	0.4887	0.5475	0.5987
32	0.1111	0.2603	0.7224	0.6158	0.6342	0.5194	0.4923	0.5462	0.5971
64	0.1140	0.2645	0.7177	0.6203	0.6379	0.5122	0.4902	0.5538	0.5980
100	0.1122	0.2614	0.7207	0.6205	0.6321	0.5061	0.4904	0.5567	0.5979

Appendix

Table 7: Performance variation according to the number of neighbors (

k

) in local attention. The best results are in bold .

$\tau_{\text{neg}}$	ST-Net			Her2ST			SCC
$\tau_{\text{neg}}$	MSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$	MSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$	MSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$
0.6	0.1111	0.2603	0.7224	0.6158	0.6342	0.5194	0.4923	0.5462	0.5971
1.0	0.1128	0.2623	0.7191	0.6240	0.6374	0.5104	0.4926	0.5497	0.5944
1.4	0.1129	0.2624	0.7212	0.6236	0.6371	0.5064	0.4966	0.5481	0.5952

Appendix

Table 8: Impact of the temperature parameter (

\tau_{\text{neg}}

) on model performance. The best results are in bold .

$\beta$	ST-Net			Her2ST			SCC
$\beta$	MSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$	MSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$	MSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$
0.75	0.1108	0.2565	0.7253	0.6229	0.6358	0.5148	0.4907	0.5572	0.6002
1.5	0.1111	0.2603	0.7224	0.6158	0.6342	0.5194	0.4923	0.5462	0.5971
3.0	0.1155	0.2652	0.7088	0.6251	0.6397	0.5113	0.4969	0.5504	0.5952

Appendix

Table 9: Impact of the negative attention scaling factor (

\beta

) on model performance. The best results are in bold .

Model	ST-Net			Her2ST			SCC
Model	MSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$	MSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$	MSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$
UNI2-h [ 4 ]	0.1111	0.2603	0.7224	0.6158	0.6342	0.5194	0.4923	0.5462	0.5971
Prov-GigaPath [ 28 ]	0.1147	0.2645	0.7125	0.6279	0.6472	0.5121	0.4914	0.5468	0.5988

Appendix

Table 10: Performance comparison between different feature extractors. The best results are in bold .

Figure 8: Additional illustration of the effect of pseudo-spots (off-grid sampling) on target spot prediction.

Figure 9: Comparison of attention maps with and without negative-aware attention.

Figure 10: Comparison of attention maps with and without negative-aware attention.

Figure 11: Bar plots of PCC scores for predicted FASN and GNAS gene expressions for each sample in the ST-Net dataset. The average PCC across the entire dataset is indicated in parentheses within the legend.

Figure 12: Histograms of PCC scores for FASN and GNAS predictions by the two methods on the ST-Net dataset. The distributions show that FEAST achieves higher PCC values compared to MERGE. Notably, it is evident that FEAST rarely yields negative PCC values.

Figure 13: Qualitative comparison of predicted FASN gene expression heatmaps.

Figure 14: Qualitative comparison of predicted GNAS gene expression heatmaps.

Spatial Gene Expression Prediction

Spatial gene expression prediction infers tissue-resolved molecular activity from histology and spatial context.

See all 38 papers

Sep 14, 2026

Hyperbolic Contrastive Learning with Entailment for Spatial Transcriptomics

Spatial Transcriptomics (ST) has transformed biomedical research by enabling the spatial mapping of gene expression across tissue sections. However, high operational costs, specialized equipment requirements, and sensitivity to experimental noise limit the accessibility and scalability of ST. Recent computer vision approaches aim to overcome these limitations by predicting spatial gene expression directly from histopathology images. While effective, current approaches often suffer from gene expression over-smoothing and overly uniform predictions across tissue regions, suggesting that further progress depends on learning representations that reflect the hierarchical and asymmetric structure of gene regulation and tissue morphology. To address these issues, we propose Hyperbolic Contrastive Learning with Entailment for Spatial Transcriptomics (HyCLoST), a hyperbolic contrastive learning model that captures the intrinsic hierarchical relationships within ST data. By leveraging hyperbolic geometry and a gene-to-image entailment loss, HyCLoST learns structured, biologically grounded representations that improve gene expression prediction accuracy, achieving a 6% reduction in MSE and an 8% increase in PCC across 26 ST datasets, over previous methods. Our source code is publicly available at https://github.com/BCV-Uniandes/HyCLoST

Daniela Vega, Paula Cárdenas, Hannah Ceballos +2

Center for Research and Formation in Artificial Intelligence Universidad de los Andes, Colombia

Contrastive Learning Spatial Transcriptomics Spatial Gene Expression Prediction Gene Expression Prediction from Histopathology Images Whole-Slide Image Analysis Hyperbolic Representation Learning

Ask PDF

May 6, 2026

HEXST: Hexagonal Shifted-Window Transformer for Spatial Transcriptomics Gene Expression Prediction

Spatial transcriptomics offers spatially resolved gene expression profiling within tissue sections, but its cost and limited throughput hinder large-scale deployment. To extend this capability to routine practice, recent computational methods aim to infer spatial gene expression directly from ubiquitous hematoxylin and eosin-stained histology slides. However, most existing models assume Cartesian or geometry-agnostic locality, despite the hexagonal sampling of widely used spot-array platforms, and point-wise regression objectives often yield over-smoothed gene expression profiles, obscuring gene-specific spatial heterogeneity. To address these, we propose HEXST, a geometry-aligned Transformer for spatial gene expression prediction from histology. HEXST operates directly on hexagonal spot coordinates to enable efficient local-to-global contextual modeling via tailored shifted-window attention mechanism and hexagonal rotary positional encoding. To enhance gene-wise spatial contrast, HEXST complements point-wise regression with a contrast-sensitive differential objective and transcriptomic priors from a pretrained single-cell foundation model during training. Across seven spatial transcriptomics datasets, HEXST consistently outperforms state-of-the-art models, providing accurate and robust spatial gene expression predictions while preserving gene-wise contrast and spatial heterogeneity.

Keunho Byeon, Jin Tae Kwak

School of Electrical Engineering, Korea University, Seoul

Computational Pathology Spatial Transcriptomics Transformer Attention Spatial Gene Expression Prediction

Ask PDF

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

Abstract

Spatial Gene Expression Prediction

Explore similar work

Hyperbolic Contrastive Learning with Entailment for Spatial Transcriptomics

HEXST: Hexagonal Shifted-Window Transformer for Spatial Transcriptomics Gene Expression Prediction

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