STRATA: Reinforcement-driven Active Sampling with Tri-paradigm Adaptation for Efficient 3D Spatial Transcriptomics

Junchao Zhu; Ruining Deng; Chongyu Qu; Zhengyi Lu; Yanfan Zhu; Yuechen Yang; Yu Wang; Shilin Zhao

cs.CVOct 29, 2026

STRATA: Reinforcement-driven Active Sampling with Tri-paradigm Adaptation for Efficient 3D Spatial Transcriptomics

Authors: Junchao Zhu, Ruining Deng, Chongyu Qu, Zhengyi Lu, Yanfan Zhu, Yuechen Yang, Yu Wang, Shilin Zhao,

Organizations: Vanderbilt University, Tennessee, USA · Weill Cornell Medicine, New York, USA · Vanderbilt University Medical Center, Tennessee, USA

Abstract

Spatial transcriptomics (ST) enables joint characterization of tissue morphology and gene expression. However, 3D volumetric ST acquisition remains expensive. AI-driven methods offer a low-cost alternative but remain limited in accuracy, while recent 2D section-wise ST acquisition still rely on dense grid sampling, resulting in redundant measurements and undersampling of biologically informative regions. To bridge the gap, we propose STRATA, a reinforcement-driven active sampling framework with tri-paradigm adaptation for reconstructing dense 3D transcriptomic maps from sparse voxel-wise ST acquisition. STRATA constructs a 3D graph integrating H&E morphology with single-cell priors and employs a reinforcement learning agent to estimate morphological sparsity and transcriptomic informativeness, enabling budget-aware selection of informative spots. Given these sparse measurements, an adaptive tri-paradigm predictor combines retrieval, regression, and cross-layer imputation through lightweight fusion, dynamically weighting their contributions. Experiments across datasets demonstrate improved sampling efficiency and accuracy, particularly under low-budget settings. Code is available at https://github.com/Junchao-Zhu/STRATA.

Figures & tables

Figure 1: Overview of STRATA. We first align serial H&E sections into a 3D voxel space and construct a cellular prototype dictionary from single-cell reference to introduce biological priors. We then develop a cell-aware RL sampling agent that evaluates candidate spots by morphological sparsity and bio-informativeness, and optimizes selection with validation loss reduction, 3D redundancy penalty, and prototype coverage bonus. Finally, an adaptive tri-paradigm predictor integrates visual retrieval, regression, and 3D cross-layer imputation via a lightweight head to generate high-fidelity expression predictions.

Method	HER2 [ 1 ]		Breast [ 4 ]		cSCC [ 5 ]		Kidney
Method	MSE $\downarrow$	PCC $\uparrow$	MSE $\downarrow$	PCC $\uparrow$	MSE $\downarrow$	PCC $\uparrow$	MSE $\downarrow$	PCC $\uparrow$
2D Monolithic Models
ST-Net [ 4 ]	0.527	0.271	0.623	0.152	1.515	0.137	0.728	0.227
His2ST [ 12 ]	0.603	0.244	0.740	0.148	1.667	0.117	0.776	0.148
EGN [ 9 ]	0.603	0.268	0.677	0.124	1.601	0.137	0.770	0.118
mclSTExp [ 6 ]	0.593	0.259	0.745	0.128	1.804	0.142	0.871	0.201

Table 1: Quantitative comparison on gene expression prediction. STRATA achieves the best overall performance.

Figure 2: Comparison of active sampling strategies , reporting PCC across backbones and datasets. STRATA achieves superior or competitive performance in most settings, with notable gains in the low-budget regime.

Figure 3: Reward ablation on the RL sampling agent. Each curve denotes one reward variant.

Single Branch
Model	HER2 [ 1 ]		cSCC [ 5 ]
Model	MSE $\downarrow$	PCC $\uparrow$	MSE $\downarrow$	PCC $\uparrow$
w.o. 3D $L_{con}$	0.434 $\pm\,0.132$	0.314 $\pm\,0.053$	1.500 $\pm\,0.083$	0.155 $\pm\,0.177$
STRATA Reg	0.404 $\pm\,0.132$	0.328 $\pm\,0.052$	1.524 $\pm\,0.080$	0.177 $\pm\,0.194$
STRATA Ret	0.443 $\pm\,0.124$	0.287 $\pm\,0.110$	1.431 $\pm\,0.063$	0.190 $\pm\,0.165$
Multiple Branch Fusion

Table 2: Ablation study in STRATA. Individual branches provide complementary signals, while adaptive fusion consistently improves performance.

Spatial Transcriptomics

Spatial transcriptomics maps gene activity to tissue locations to reveal cell states, neighborhoods, and disease-related organization.

See all 64 papers

Benchmarking Active Spot Selection for Cost-Efficient Spatial Transcriptomics
Sep 23, 2026Zheyu Zhu, Junchao Zhu, Fengbei Liu +7Spatial Transcriptomics Active Learning
University of Pennsylvania, Philadelphia, PA 19104, USA · Vanderbilt University, Nashville, TN, USA · Cornell Tech, New York, NY 10044, USA +4
HiST: A Hierarchical Sparse Transformer for Cross-Modal Spatial Transcriptomics Modeling
Jun 12, 2026Weiyi Wu, Xinwen Xu, Xingjian Diao +4Computational Pathology Spatial Transcriptomics
Dartmouth College · Mass General Hospital · New Jersey Institute of Technology +1
Spatial Transcriptomics as Images for Large-Scale Pretraining
Mar 13, 2026Yishun Zhu, Jiaxin Qi, Jian Wang +2Spatial Transcriptomics
Computer Network Information Center, Chinese Academy of Sciences, Beijing, China · Hangzhou Institute for Advanced Study, University of the Chinese Academy of Sciences, Zhejiang, China

Sep 23, 2026J/K move · Enter open · S save

Benchmarking Active Spot Selection for Cost-Efficient Spatial Transcriptomics

Zheyu Zhu, Junchao Zhu, Fengbei Liu, Tianyuan Yao+6

University of Pennsylvania, Philadelphia, PA 19104, USA · Vanderbilt University, Nashville, TN, USA · Cornell Tech, New York, NY 10044, USA+4

Spatial Transcriptomics Active Learning Training Data Selection Uncertainty Sampling Spatial Gene Expression Prediction

Open paper AskPDF arXiv

Spatial transcriptomics (ST) measures gene expression in tissue context, but dense capture grids can be costly and may repeatedly sample morphologically similar regions. Most active learning strategies were developed for categorical labels and independent samples. We conduct a retrospective pool-based benchmark of active learning versus uniform Random sampling for ST, where expression vectors are high-dimensional and continuous and candidates are spatially correlated. Using two fully profiled public ST cohorts, we mask candidate expression vectors and simulate multi-round selection with uncertainty-based Monte Carlo dropout (MC-dropout) and temporal output discrepancy (TOD), and diversity-based CoreSet and TypiClust-inspired selection. We compare 160 completed configurations at 5%, 10%, 30%, and 50% of the fold-wide training spot pool under patient-level cross-validation, with a separate full-label reference. Within each budget, strategies share the selection schedule, morphology-to-expression predictor, and optimization protocol. We assess mean per-gene within-slide Pearson correlation coefficient (PCC), expression-cluster agreement, and Moran's I fidelity. On HER2-positive breast cancer, pooled mean PCC differences from Random across the four active strategies were -0.0176, -0.0117, +0.0056, and +0.0057 at 5%, 10%, 30%, and 50%, respectively. On cutaneous squamous cell carcinoma (cSCC), three strategies were below Random at 5%, and all four were below Random at 10%. On HER2-positive breast cancer, CoreSet and MC-dropout had lower PCC but higher expression-cluster agreement than Random at the two smallest budgets; this pattern did not reproduce on cSCC. Under the reported fixed training horizons, the evaluated active strategies do not consistently improve on Random at small budgets, and rankings depend on the evaluation measure.

Abstract

Spatial Transcriptomics

Explore similar work

Benchmarking Active Spot Selection for Cost-Efficient Spatial Transcriptomics

HiST: A Hierarchical Sparse Transformer for Cross-Modal Spatial Transcriptomics Modeling

Spatial Transcriptomics as Images for Large-Scale Pretraining