Reinforcement Learning and SCR2-ST Unlock Efficient Spatial Transcriptomics Data Acquisition

Researchers are increasingly focused on understanding how gene expression relates to tissue structure, but acquiring sufficient data using spatial transcriptomics remains a significant challenge due to its cost. Junchao Zhu, Ruining Deng, and Junlin Guo, alongside colleagues, address this limitation with SCR2-ST, a novel framework that intelligently combines single-cell sequencing data with spatial transcriptomics.

The team’s approach uses reinforcement learning to guide the selection of informative tissue regions for sequencing, dramatically improving sampling efficiency and prediction accuracy, particularly when resources are limited. By integrating prior biological knowledge from single-cell data and employing a hybrid prediction network, SCR2-ST represents a substantial advance in the field, offering a powerful new tool for exploring the complexities of tissue biology. Recognizing that acquiring ST data remains expensive and traditional fixed-grid sampling often captures redundant information, the team engineered a system that selectively samples informative tissue regions, maximizing biological diversity within constrained sequencing budgets. This approach leverages the wealth of data available from single-cell sequencing, which typically exceeds ST data by an order of magnitude, to guide the sampling process and improve prediction accuracy. The core of SCR2-ST is a single-cell guided reinforcement learning (SCRL) active sampling method, which combines single-cell foundation model embeddings with spatial density information to construct biologically grounded reward signals. This innovative system dynamically selects spots for sequencing, prioritizing regions that offer the most informative biological content, rather than relying on a pre-defined grid. SCRL effectively avoids redundant measurements and prioritizes regions with high biological diversity.

The team then developed SCR2Net, a hybrid regression-retrieval prediction network that leverages the actively sampled data. SCR2Net combines regression-based modeling with retrieval-augmented inference, effectively predicting gene expression profiles based on both local image appearance and comparisons to similar profiles in the single-cell data. To further refine prediction accuracy, scientists incorporated a majority cell-type filtering mechanism within SCR2Net, suppressing noisy matches and ensuring that retrieved expression profiles serve as reliable “soft labels” for auxiliary supervision. The framework was rigorously evaluated on three public ST datasets, demonstrating state-of-the-art performance in both sampling efficiency and prediction accuracy, particularly under low-budget scenarios. This breakthrough delivers a powerful new approach to spatial transcriptomics, enabling more efficient and accurate mapping of gene expression within tissues and opening new avenues for biological discovery.

Smart Spatial Sampling with Reinforcement Learning Scientists have developed a new framework, SCR2-ST, that integrates existing single-cell data with spatial transcriptomics to improve the efficiency of data acquisition and the accuracy of gene expression prediction. Traditional spatial transcriptomics methods often acquire data from many redundant or uninformative regions, which is costly and limits analysis, but this new approach actively selects which tissue regions to sample, focusing on those most likely to yield valuable biological insights. The framework achieves this through two key components: a reinforcement learning-based sampling strategy and a hybrid prediction network. The sampling strategy uses prior knowledge from single-cell data alongside spatial density information to guide the selection of informative regions, avoiding unnecessary measurements. The prediction network then combines direct regression modelling with retrieval-augmented inference, using cell-type filtering to reduce noise and improve the accuracy of gene expression predictions, particularly when data acquisition is limited by budget constraints. Testing on multiple datasets demonstrates that SCR2-ST outperforms existing methods in both sampling efficiency and prediction accuracy, especially under low-budget scenarios. 👉 More information 🗞 SCR2-ST: Combine Single Cell with Spatial Transcriptomics for Efficient Active Sampling via Reinforcement Learning 🧠 ArXiv: https://arxiv.org/abs/2512.13635 Tags: Rohail T. As a quantum scientist exploring the frontiers of physics and technology. My work focuses on uncovering how quantum mechanics, computing, and emerging technologies are transforming our understanding of reality. I share research-driven insights that make complex ideas in quantum science clear, engaging, and relevant to the modern world. Latest Posts by Rohail T.: Chaplygin Gas and Potential Model Advances Understanding of Early Universe Resonances December 16, 2025 Quantum NIZK Proofs Advance Certified-Everlasting Zero-Knowledge with Statistical Indistinguishability December 16, 2025 Quantum Chemistry Achieves 0.94 Accuracy, Paving Way for Quantum Advantage December 16, 2025