TransMap
Image-native representations for single-cell genomics via Gromov–Wasserstein optimal transport.
TransMap reorganises sparse single-cell gene-expression vectors as 2D feature images by placing genes on a fixed pixel grid via discrete optimal transport, so that nearby pixels carry related genes. Modern convolutional architectures can then exploit gene–gene structure as spatial locality — without the parameter cost of token-based foundation models.
Method. A Gromov–Wasserstein coupling between a gene–gene distance matrix (STRING co-expression evidence or pseudobulk Pearson) and a fixed grid metric is solved with entropic regularisation; the soft transport plan is rounded to an injective hard assignment via sparse bipartite matching. For cross-species analysis, a fused Gromov–Wasserstein BCD procedure aligns multiple species on a shared grid, treating ortholog pairs as harmonic-weighted feature costs rather than collapsing them into a single feature axis.
Results. A ~15M-parameter CNN on TransMap images beats parameter-matched MLPs (scVI) on five multi-organ scRNA-seq benchmarks (up to 714k cells / 167 donors) on scIB embedding and scGraph structural metrics. A shared 5-species grid aligns human–mouse pancreas cells (iLISI = 1.297, 76.4% label-transfer accuracy) without ortholog conversion. A random-layout ablation confirms the gains come from the geometry, not the architecture: random layouts plateau at ~655 val loss versus ~462 for GW layouts.
Code release planned alongside publication.