Spatial transcriptomics provides spatially resolved molecular profiling, but its cost, limited throughput and experimental complexity hinder routine large-scale use. Predicting spatial gene expression from hematoxylin-and-eosin-stained histology images offers a scalable alternative, yet remains challenging because histology provides only an indirect morphological view of molecular activity. Existing methods typically decode expression from visual features in a single stage, even when enhanced with spatial aggregation, retrieval or multi-resolution fusion. This formulation is restrictive because spot-level expression reflects not only local morphology but also cellular identity, neighborhood composition and tissue-level organization. We propose BCARP, a Biology-guided Context-Aware Residual Propagation framework for histology-guided spatial transcriptomics prediction. Instead of directly predicting gene expression from histology embeddings, BCARP treats them as morphology visual priors and progressively refines spot representations in a transcriptomics-oriented latent space. Each refinement step integrates nuclei-guided local cellular context with prototype-guided global context propagation, enabling information exchange across both neighboring and spatially distant spots. A pattern-scale decoder further separates spatial expression into pattern prediction from raw expression with scale calibration. Evaluated on three ST datasets and ten HEST-1k cohorts, BCARP achieves the highest mean correlation-based performance against other baselines. Numerical results support biology-guided local-global residual propagation as an effective alternative to visual decoding for histology-to-transcriptomics prediction.
Figures & tables
Figure 1: Pipeline overview of BCARP. Frozen histology foundation embeddings are first projected into a shared latent space, then iteratively refined through context-conditioned modulation, local biological encoding and prototype-guided global context propagation. The final representation is passed to a pattern-scale gene expression decoder to generate calibrated spatial gene expression predictions.
Figure 2: ECDF curves show the distribution of fold-averaged per-gene PCC on HER2ST, Skin and ST-Net datasets. Dashed vertical lines indicate the average per-gene PCC of BCARP and the highest competing baseline.
HER2ST
Skin
ST-Net
Model
PCC-H ↑
PCC-M ↑
MSE ↓
MAE ↓
RVD ↓
PCC-H ↑
PCC-M ↑
MSE ↓
MAE ↓
RVD ↓
PCC-H ↑
PCC-M ↑
MSE ↓
MAE ↓
RVD ↓
Frozen-feature references
ResNet50+Ridge
0.2249
0.2291
1.0782
0.8224
0.9039
0.2695
0.3260
1.3565
0.9050
0.8581
0.0985
0.0890
0.9762
0.7862
0.9125
UNI+Ridge
0.2583
0.2598
1.0579
0.8144
0.8840
0.2895
0.3262
1.3382
0.8968
0.8812
0.1096
0.0936
0.9666
0.7821
0.9223
UNI+MLP
0.3474
0.3316
1.0030
0.7890
0.8452
0.3029
0.3604
1.2547
0.8670
0.8368
0.1843
0.1633
0.9195
0.7609
0.8644
UNI+Decoder
0.4249
0.3721
0.9431
0.7724
0.8188
0.5392
0.4930
1.4455
0.9538
1.4511
0.2319
0.1868
0.8283
0.7351
0.7685
Table 1: Performance comparison on HER2ST, Skin and ST-Net datasets. PCC-H and PCC-M denote average per-gene PCC over highly variable genes and all target genes, respectively. MSE and MAE measure prediction error in expression space, and RVD measures the relative deviation in expression variance between predictions and ground truth. Among the method baselines, best results are shown in bold and second-best results are underlined; the frozen-feature references are reported for context and are excluded from this ranking.
Model
SKCM
LUNG
PAAD
READ
COAD
LYMPH
CCRCC
IDC
PRAD
HCC
Mean
Mean Rank ↓
Frozen-feature ridge probes
ResNet50+Ridge
0.5561
0.5310
0.3963
0.0738
0.3463
0.2642
0.1729
0.5564
0.2005
0.0676
0.3165
10.8
CONCH+Ridge
0.6138
0.5819
0.4585
0.1941
0.4214
0.3131
0.1816
0.5995
0.3031
0.0780
0.3745
7.7
UNI+Ridge
0.6340
0.5833
0.4863
0.1987
0.4289
0.3020
0.2077
0.6212
0.2308
0.0868
0.3780
6.3
GigaPath+Ridge
0.6025
0.5876
0.4823
0.2084
0.4324
0.2958
0.1967
0.6122
0.2886
0.0948
0.3801
6.2
Virchow2+Ridge
0.6334
0.6005
0.4805
0.2051
0.4220
0.3029
0.2250
0.6411
0.2753
0.1118
0.3898
5.4
Table 2: PCC-M performance across the ten HEST-1k cohorts. Mean denotes the average PCC-M across cohorts, and Mean Rank denotes the average within-cohort rank, where lower is better. Best results are shown in bold, and second-best results are underlined.
Figure 3: Qualitative comparison on HER2ST for FN1 on 3 slides.
Variant
PCC-H ↑
PCC-M ↑
MSE ↓
MAE ↓
RVD ↓
Full
0.4567
0.4032
0.8381
0.7191
0.7084
Main components
w/o Proj.
0.4283
0.3772
0.9029
0.7512
0.8592
w/o CMod
0.4379
0.3864
0.9608
0.7751
0.7870
w/o Local
0.4421
0.3882
0.8873
0.7498
0.8138
w/o Global
0.4330
0.3799
0.9264
0.7646
0.8092
Table 3: Component ablation on HER2ST.
Figure 4: Relative percentage change of each metric under different refinement depths compared with R=1 on ST-Net dataset.
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Spatial Transcriptomics Prediction from Histology
Figure 6: List of the 250 genes used in all experiments on the three ST datasets.
HER2ST [ 1 ]
Skin [ 18 ]
ST-Net [ 12 ]
Model
PCC-H ↑
PCC-M ↑
MSE ↓
MAE ↓
RVD ↓
PCC-H ↑
PCC-M ↑
MSE ↓
MAE ↓
RVD ↓
PCC-H ↑
PCC-M ↑
MSE ↓
MAE ↓
RVD ↓
Frozen-feature reference models
ResNet50+Ridge
0.2249
0.2291
1.0782
0.8224
0.9039
0.2695
0.3260
1.3565
0.9050
0.8581
0.0985
0.0890
0.9762
0.7862
0.9125
UNI+Ridge
0.2583
0.2598
1.0579
0.8144
0.8840
0.2895
0.3262
1.3382
0.8968
0.8812
0.1096
0.0936
0.9666
0.7821
0.9223
UNI+MLP
0.3474
0.3316
1.0030
0.7890
0.8452
0.3029
0.3604
1.2547
0.8670
0.8368
0.1843
0.1633
0.9195
0.7609
0.8644
UNI+Decoder
0.4249
0.3721
0.9431
0.7724
0.8188
0.5392
0.4930
1.4455
0.9538
1.4511
0.2319
0.1868
0.8283
0.7351
0.7685
Appendix
Table 4: Performance comparison across the three ST datasets under their native architectural and encoder configurations. Bold indicates the best result and underlining indicates the second-best result; frozen-feature references are excluded from this ranking.
Foundation Backbone
Pathology foundation encoder: UNI [ 3 ]
Spot patch size: 224
Embedding dimension de : 1024
Refinement iterations R : 5
Biology-guided Context Modeling
Spatial neighbors k : 8
Appendix
Table 5: Hyperparameter configuration of BCARP.
Labels
PCC-H ↑
PCC-M ↑
MSE ↓
MAE ↓
RVD ↓
HoVer-Net [ 9 ]
0.4567
0.4032
0.8381
0.7191
0.7084
CellViT [ 14 ]
0.4634
0.4060
0.8769
0.7426
0.6998
Appendix
Table 6: Impact of the nuclei-analysis backend on HER2ST. All other settings are identical; only the pseudo-label generator changes.
Method / component
Trainable (M)
E2E / slide (s)
Peak (GB)
Baselines based on UNI embedding
UNI+Ridge
0.26
1.28
0.22
mclSTExp [ 23 ]
41.46
1.29
0.75
MERGE [ 8 ]
21.53
1.29
0.84
STFlow [ 15 ]
1.35
1.32
1.03
TRIPLEX [ 5 ]
63.12
1.57
6.88
Appendix
Table 7: Computational cost on a single RTX 5070 Ti.
Figure 7: Distribution of fold-averaged per-gene Pearson correlation coefficients (PCC-H and PCC-M) across HER2ST [ 1 ] , ST-Net [ 12 ] and Skin [ 18 ] . BCARP generally shifts the per-gene correlation distributions toward higher values across both the full target panel and the HVG subset.
Figure 8: Per-fold PCC-M performance across the three ST datasets.
Figure 9: Per-fold PCC-M performance across the ten HEST-1k cohorts. Dashed lines mark each method’s cohort mean.
Dataset
Metric
1
2
3
4
5
6
7
8
9
10
HER2ST [ 1 ]
PCC-H ↑
+0.0
+5.4
+6.3
+8.1
+7.9
+7.9
+7.4
+8.5
+8.1
+7.6
PCC-M ↑
+0.0
+5.7
+6.1
+7.8
+7.5
+7.5
+7.4
+8.4
+7.7
+7.4
MSE ↓
+0.0
-6.8
-2.0
-5.1
-5.7
-2.3
-2.5
-7.4
-3.9
-3.8
MAE ↓
+0.0
-4.2
-0.7
-1.9
-2.5
-1.0
-1.4
-3.7
-1.6
-1.7
RVD ↓
+0.0
-14.3
+18.4
-12.6
+1.7
+7.1
+11.6
-11.7
+7.9
+10.2
Skin [ 18 ]
PCC-H ↑
+0.0
+1.2
+1.2
+1.6
+1.0
+1.1
+1.5
+1.8
+1.1
+1.1
Appendix
Table 8: Refinement-depth ablation as relative change (%) of each metric with respect to the anchor R=1 .
Kctx
PCC-H ↑
PCC-M ↑
MSE ↓
MAE ↓
RVD ↓
Mean Imp. ↑
0
0.4586
0.4061
0.8538
0.7282
0.8123
0.00%
5
0.4608 ( + 0.5%)
0.4076 ( + 0.4%)
0.8933 ( + 4.6%)
0.7507 ( + 3.1%)
0.7522 ( − 7.4%)
+0.11%
10
0.4509 ( − 1.7%)
0.3941 ( − 3.0%)
0.8807 ( + 3.2%)
0.7435 ( + 2.1%)
0.9480 ( + 16.7%)
− 5.32%
25
0.4534 ( − 1.1%)
0.4043 ( − 0.4%)
0.9277 ( + 8.7%)
0.7657 ( + 5.1%)
0.8649 ( + 6.5%)
− 4.37%
50
0.4567 ( − 0.4%)
0.4032 ( − 0.7%)
0.8381 ( − 1.8%)
0.7191 ( − 1.2%)
0.7084 ( − 12.8%)
+2.95%
100
0.4374 ( − 4.6%)
0.3891 ( − 4.2%)
0.8903 ( + 4.3%)
0.7537 ( + 3.5%)
0.7987 ( − 1.7%)
− 2.98%
Appendix
Table 9: Sensitivity to the number of global prototype tokens Kctx on HER2ST. Mean Imp. denotes the mean sign-corrected relative improvement across the five metrics, where decreases in MSE, MAE, and RVD are counted as positive improvements.
Dataset
Fold
Test partition
Train slides
Test slides
Train spots / Test spots
ST datasets
HER2ST [ 1 ]
0
A
30
6
11,555 / 2,065
1
B
30
6
11,908 / 1,712
2
C
30
6
12,534 / 1,086
3
D
30
6
11,787 / 1,833
4
E
33
3
11,891 / 1,729
Appendix
Table 10: Detailed cross-validation splits by fold. For HER2ST, Skin and ST-Net, test partitions are identified by patient; all sections from a patient are assigned exclusively to the training or the test partition. For the HEST-1k cohorts we adopt the official patient-grouped folds released with the benchmark and identify test partitions by HEST slide ID.
Figure 10: Qualitative comparison of predicted expression for three marker genes on Skin sample P2_rep3.
Figure 11: Qualitative comparison of predicted expression for three marker genes on Skin sample P5_rep2.
Figure 12: Qualitative comparison of predicted expression for three marker genes on Skin sample P10_rep2
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
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