Spatial Transcriptomics as Images for Large-Scale Pretraining

A team of researchers proposes recasting spatial transcriptomics data as croppable, multi-channel images to enable large-scale pretraining of deep learning models, addressing a fundamental bottleneck in how tissue-level gene expression is organized for machine learning [1]. Spatial transcriptomics, or ST, measures thousands of gene expression values at discrete spots with precise coordinates on tissue sections, preserving spatial context that is essential for clinical and pathological studies [2]. As sequencing throughput and platform capabilities advance, the expanding data volumes have motivated efforts toward large-scale ST pretraining. But the field has lacked consensus on what constitutes a single training sample [1]. Existing approaches fall into two camps, each with drawbacks. Spot-based methods treat every measurement location as an independent sample, which discards spatial dependencies and effectively collapses ST into single-cell transcriptomics [3]. Slice-based methods treat an entire tissue slice as one sample, preserving global spatial organization but producing prohibitively large inputs and drastically fewer training examples, undermining effective pretraining [5]. To bridge this gap, the authors define a multi-channel image representation with a fixed spatial size by cropping patches from raw slides. This preserves local spatial context while substantially increasing the number of training samples [2]. Along the channel dimension, they introduce gene subset selection rules to control input dimensionality and improve pretraining stability [4]. The approach draws on deep learning techniques that use multilayered neural networks for tasks such as classification and representation learning, architectures that have been applied across computer vision, bioinformatics, and medical image analysis [6]. In experiments, the image-like dataset construction consistently improved downstream performance compared with conventional pretraining schemes [1]. Ablation studies verified that both the spatial patching and the channel design are necessary for the gains, establishing what the authors describe as a unified, practical paradigm for organizing ST data [3]. The work was submitted to arXiv on 13 March 2026 and revised on 3 June 2026 [1].