STING: A Graph Neural Network Approach for Computational Inference of Spatial Transcriptomic Profiles

Abstract

Spatial transcriptomics enables the measurement of mRNA counts at spatial locations within a tissue but faces challenges such as high experimental costs, technical expertise requirements, and low RNA detection efficiency at high resolution. We present STING (Spatial Transcriptomics Inference using Graph neural networks), a computational approach that infers spatial gene expression patterns with high efficiency. STING integrates convolutional neural networks (CNNs) pre-trained on histological images with graph neural networks (GNNs) to model spatial proximity, representing tissue sections as nearest-neighbor graphs where spatially adjacent spots are interconnected. We train GNNs on these graphs to predict gene expression at additional tissue locations, including unseen samples. Evaluated on two public spatial transcriptomics datasets (59 and 36 tissues), STING achieves a Pearson correlation coefficient (PCC) of up to 0.79 for super-resolution and 0.69 for tissue-wide inference, outperforming existing methods in accuracy. Our results demonstrate that STING is an effective and computationally efficient tool for predicting spatial gene expression, with significant applications in cancer research, personalized medicine, and beyond.