Tutorials

Tutorials#

Explanatory notebooks organized by FuseMap’s core capabilities. Integration and mapping examples cover data preparation, running FuseMap, and analysis. Tutorials 3.1, 3.3, and 3.4 reuse Tutorial 1.2’s outputs; precomputed outputs are available for Tutorials 1.2, 1.3, and 1.4. Tutorial 2.3 requires data available upon request. The Colab setup cells prepare Python 3.11 when needed and restart the session; reconnect and rerun the setup cell before continuing. The runtime must have enough memory for the selected dataset. See Installation for local setup and Example data for download requirements.

Integrate spatial atlases#

FuseMap trains one encoder per section against shared cell (\(Z_c\)) and tissue (\(Z_T\)) embedding spaces, so sections with different gene panels, technologies, and conditions become jointly analyzable. Sequencing-based bead/spot datasets (Slide-seq, Stereo-seq, Visium HD) are deconvolved automatically when declared as bead data. After integration, annotations transfer across datasets through the shared embedding, and cell types / tissue regions can be redefined de novo from it.

1.1 Integrate imaging-based data across platforms

1.2 Integrate imaging- and sequencing-based data (with deconvolution)

1.3 Integrate across conditions (healthy × disease)

1.4 Beyond the brain: whole-embryo cross-resolution integration

Map new data to a reference#

Project new sections onto a pretrained FuseMap model and transfer annotations — either a model you trained yourself (like the ones from the integration tutorials) or the molCCF mouse brain atlas. Mapping works best against large, diverse references such as molCCF; for small custom references, consider integrating jointly instead (see the note in Tutorial 2.1).

2.1 Map to a customized pretrained model

2.2 Map to the molCCF mouse brain atlas

2.3 Cross-species: map marmoset onto the mouse molCCF

Downstream analysis#

Everything here runs on embeddings a FuseMap run has already produced. The universal gene embedding ties every panel to a shared gene space — so FuseMap can impute genes that were never measured in a section (below, genes absent from the STARmap panel are imputed from the sequencing-based datasets integrated in Tutorial 1.2), and, because the embedding groups genes by function, it can also guide the design of a compact targeted panel. The cell and tissue embeddings drive spatial interaction analysis and reveal structure finer than any single annotation.