MagNet: Multi-Level Attention Graph Network for Predicting High-Resolution Spatial Transcriptomics

TL;DR

MagNet tackles the challenge of predicting high-resolution spatial transcriptomics by fusing multi-resolution image patch features through cross-attention and enriching spatial context with a GAT-Transformer block. It introduces a unified framework for cross-resolution feature aggregation, a spatially guided graph module, and a hybrid loss that enforces cross-scale consistency, enabling HD predictions at 8 μm and 16 μm scales. Evaluations on private kidney and public CRC HD ST datasets show state-of-the-art performance across resolutions, with ablations confirming the contributions of multi-resolution fusion, graph-based neighborhood modeling, and consistency constraints. The work provides a concrete methodology and benchmarks for high-resolution ST prediction and is released as open-source for reproducibility and further development.

Abstract

The rapid development of spatial transcriptomics (ST) offers new opportunities to explore the gene expression patterns within the spatial microenvironment. Current research integrates pathological images to infer gene expression, addressing the high costs and time-consuming processes to generate spatial transcriptomics data. However, as spatial transcriptomics resolution continues to improve, existing methods remain primarily focused on gene expression prediction at low-resolution spot levels. These methods face significant challenges, especially the information bottleneck, when they are applied to high-resolution HD data. To bridge this gap, this paper introduces MagNet, a multi-level attention graph network designed for accurate prediction of high-resolution HD data. MagNet employs cross-attention layers to integrate features from multi-resolution image patches hierarchically and utilizes a GAT-Transformer module to aggregate neighborhood information. By integrating multilevel features, MagNet overcomes the limitations posed by low-resolution inputs in predicting high-resolution gene expression. We systematically evaluated MagNet and existing ST prediction models on both a private spatial transcriptomics dataset and a public dataset at three different resolution levels. The results demonstrate that MagNet achieves state-of-the-art performance at both spot level and high-resolution bin levels, providing a novel methodology and benchmark for future research and applications in high-resolution HD-level spatial transcriptomics. Code is available at https://github.com/Junchao-Zhu/MagNet.

Figures (4)

• Figure 1: Spatial transcriptomics data at different resolutions. (A) Traditional low-resolution 10X Visium v2 barcoded spots, where spots are discretely distributed with a diameter of 55 $\mu$m. (B) Current high-resolution 10X Visium HD barcoded squares, where bins are densely distributed with a diameter of 8 $\mu$m.
• Figure 2: The network structure of the proposed MagNet. MagNet utilizes cross-attention layers to integrate features extracted from multi-resolution patches. Additionally, it incorporates a GAT-Transformer block to aggregate neighborhood information while leveraging spatial relationships. The predictions for each resolution level are then independently generated by a regression head.
• Figure 3: Qualitative comparison for pivotal gene expression prediction.
• Figure 4: Gene selection in each dataset and resolution.