PRETRAINED SPATIAL RECONSTRUCTION
Pretrain once.
Reconstruct new sections.
A pretrained model for spatial transcriptomic super-resolution. ASTRA combines tissue histology with measured RNA to reconstruct expression at finer spatial resolution, without fitting the model to each new section.
CODE & DOCUMENTATION
jiachenye-cuilab /
Implementation, setup and usage.
github.com/
THE APPROACH
ASTRA brings pretraining to spatial reconstruction and treats capture geometry as part of the input. The same checkpoint can work with dense bins and separated capture spots.
PRETRAINING
ASTRA separates source pretraining from reconstruction of a new section. Once pretrained, the model uses fixed weights for inference, removing target-section training and test-time optimization from the reconstruction workflow.
CAPTURE GEOMETRY
A parent is the region over which an RNA count was measured. During pretraining, high-resolution measurements are grouped into parents of varied shapes and sizes. This common representation allows one model to accommodate different sampling patterns.
MULTIMODAL CONTEXT
Image features at multiple scales describe tissue morphology. ASTRA combines these features with local RNA measurements and capture geometry to guide fine-scale reconstruction. Gene availability is encoded explicitly, keeping unmeasured genes distinct from measured zeros.
EXPRESSION ALLOCATION
Within each parent, expression allocation is normalized to the measured gene counts before whole-section assembly. For unmeasured gaps between capture regions, the model uses local molecular context and tissue morphology to estimate expression.
HOW IT WORKS
Prepare the aligned image and RNA inputs, apply the pretrained model, and assemble the predictions into a finer spatial map.
INPUT
Provide a registered H&E image, measured RNA counts and the spatial support of each capture region.
PRETRAINED MODEL
The model combines morphology and local molecular context to predict how expression is distributed across space.
The reconstruction weights stay fixed.
OUTPUT
Combine predictions from overlapping tissue fields into expression estimates on a finer spatial grid.
Find the source code, setup instructions and usage documentation on GitHub.