Tumor clones get positions on the slide from their family tree: SpOTr aligns phylogeny with spot distances and leads clone-proportion recovery in simulations
Synopsis
Spatial transcriptomics measures spots about 55 micrometres across that mix several cells, while dissociated single-cell data resolve clones but lose their positions. SpOTr couples spot deconvolution with Fused Gromov–Wasserstein optimal transport, using distances in the cancer cells' evolutionary tree to constrain which spots they land in, so closely related cells are drawn to nearby spots. In simulated tumors it recovered clone proportions better than cell2location, DestVI, Tangram, scDOT and expression-only transport in every scenario, and on a PEtracer lineage-traced section its clade composition agreed best with the in situ readout.
Interpretation
Cancer clones in a tumor section can now be placed using their evolutionary relationships, not only differences in gene expression. Earlier integration methods deconvolved spots into reference cell types or mapped individual cells to spots by expression, assuming the target populations have distinguishable expression profiles; tumor subclones only partly inherit their phenotypes, so that assumption fails. On simulated tumors whose lineages came from the Cassiopeia birth–death process and whose nests were laid out with scCube, SpOTr recovered spot clone composition better than cell2location, DestVI, Tangram, scDOT, exact expression-only optimal transport and FGW with the same distance matrices in every scenario; its advantage was largest when clones were hardest to distinguish and it stayed most accurate when they differed clearly.
The method optimizes spot deconvolution and the cell-to-spot mapping together in alternation. Deconvolution follows the cell2location negative binomial model, splitting each spot's counts into microenvironment cell types and cancer cells, with clone abundance set by the mapping. The mapping is an entropic optimal transport problem whose cost combines the negative log-likelihood of the spots under the deconvolution model, the cosine distance between cancer-cell and spot expression profiles, and a structural term matching tree distances between cells to spatial distances between spots; the mapping is updated by Sinkhorn iterations and the deconvolution parameters by gradient descent to a maximum a posteriori estimate.
On real data the structural term yields subclone divisions that expression-only deconvolution does not. On PEtracer section M3-4 only 31% of tumor cells carry a lineage readout, and methods were scored on clade proportions per spot; on HTAN colorectal liver metastasis case HT260C1, cell2location served as the baseline. On the PEtracer section SpOTr agreed best with the in situ truth on all four metrics clade by clade, whereas DestVI assigned most tumor spots to a single clade and Tangram and scDOT spread clades over the whole tumor; in HT260C1 SpOTr separated B1 from B2 more clearly than cell2location and placed B1a and B1b, which were not given as labels, in agreement with the CNA signal of the spots.
Perspective
The demonstrated scope covers simulated tumors, synthetic Visium spots built from a PEtracer lineage-traced section, and the HTAN colorectal liver metastasis case HT260C1; HT260C1 has no lineage readout, so its subclone placements are checked against the CNA signal of held-out genes. The intended readers are tumor and computational-methods researchers aligning dissociated single-cell data with spot-based spatial transcriptomics.
The method assumes each clone has the same proportion in the dissociated sample as in the section, which may not hold when the two samples come from different tumor regions; the structural weight alpha is set by the user rather than learned, and when clones are dispersed over the section a high alpha pulls their cells together incorrectly, with zeroing those clones' weights improving their maps in simulations. The authors propose learning per-clone weights from data and propagating tree uncertainty as extensions.
