GRAPH ATTENTION NETWORK GENERATES SUPER-RESOLUTION SPATIAL TRANSCRIPTOMIC DATA

ICLR 2026 Conference Submission16856 Authors

Keywords: Graph, spatial transcriptomics, resolution, attention, distribution shift.

Abstract: Spatial transcriptomic technologies allow for uncovering the spatial origin of RNA molecules within a tissue slide. Still, some challenges remain unsolved when acquiring informative signal. An existing trade-off hinders the choice of which one to use: sequencing-based technologies provide high-throughput profiles, while imaging-based outperform regarding spatial resolution. On the sequencing-based side, the minimal spatial unit, called spot, comprises more than one cell, yielding slightly blurred expression profiles. To avoid inaccurate analysis and misinterpretation of spatial data, we believe that cells inside a single spot should be isolated and allocated into subspots. We propose a computational method based on graphs and attention learning, named Square, that leverages message passing for information sharing between neighbor spots. Even though this rearrangement of cells can be solely spatially approximated, a resolution enhancement is achieved. We show that the proposed approach is capable of deciphering the composition of ST spots, whilst imputing sparse profiles and amplifying the signal in them. Newly generated subspots have been empirically and biologically validated. The gap between both spatial transcriptomic modalities is then closed, generating high-throughput cellular-scale outputs.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 16856