GEARS

Single-cell RNA sequencing (scRNA-seq) profiles gene expression at high resolution but discards the spatial context in which cells reside, while spatial transcriptomics (ST) preserves tissue geometry but typically measures expression at lower cellular resolution or aggregates multiple cells per measured spot. Bridging these two modalities — recovering where individual cells sit in space from their expression profiles — is a central problem for understanding tissue organization, cell-cell communication, and the microenvironment of disease. GEARS (GEometry-Aware Reconstruction of Single cells) tackles this as a generative reconstruction task, producing 2D spatial coordinates for dissociated single cells guided by reference ST data.

Introduced in a May 2026 preprint accepted at KDD 2026, GEARS is described by its authors as a "geometry-first" framework: rather than assigning cells to spots, classifying cell types, or relying on paired histology images, it directly reconstructs an intrinsic single-cell spatial geometry. The model is label-free, requiring neither cell-type annotations nor histological images nor explicit cell-to-spot assignment, which makes it applicable to settings where such auxiliary information is unavailable or unreliable. It was developed by Ehtesamul Azim, Muhtasim Noor Alif, Tae Hyun Hwang, Yanjie Fu, and Wei Zhang (corresponding) across the University of Central Florida, Vanderbilt University Medical Center, and Arizona State University.

Note that this model shares its acronym with an unrelated, earlier GEARS — the Stanford/SNAP graph-based perturbation-response predictor (Roohani et al., 2023). The two are distinct: this entry is the geometry-aware spatial reconstruction model.

#Key Features

• Geometry-first reconstruction: Directly recovers intrinsic 2D single-cell coordinates rather than assigning cells to ST spots, framing spatial inference as a distance-geometry problem solved end-to-end.
• Label-free integration: Operates without cell-type labels, histological images, or cell-to-spot assignment, broadening applicability to datasets lacking such annotations.
• Domain-invariant expression encoder: Learns representations that transfer between the scRNA-seq and ST modalities, mitigating the technical and distributional gap between dissociated and spatially resolved measurements.
• Permutation-equivariant generator: Treats cells as an unordered set so reconstructed geometry is invariant to input ordering, a natural inductive bias for dissociated single-cell populations.
• Diffusion-based refinement: Uses diffusion with EDM-style preconditioning to iteratively refine generated coordinates, improving the fidelity of the reconstructed spatial layout.

#Technical Details

GEARS combines three components into a single generative pipeline. A domain-invariant expression encoder maps both scRNA-seq and spatial transcriptomics profiles into a shared latent space, aligning the modalities so that geometry learned from ST can guide the placement of dissociated cells. A permutation-equivariant generator then proposes candidate 2D coordinates for the cell set, and a diffusion-based refinement stage — using EDM-style (Elucidated Diffusion Model) preconditioning — iteratively sharpens these coordinates. The reconstruction is cast as a distance-geometry problem, optimizing the pairwise distance structure among cells so that the recovered layout respects both global and local spatial relationships.

The authors evaluate GEARS on metrics targeting three complementary aspects of spatial fidelity: global distance preservation, local neighborhood fidelity, and spatial distribution alignment, reporting improvements over baseline methods on each. They additionally test cross-section generalization, assessing whether geometry learned on one tissue section transfers to held-out sections. As of the preprint, no public code repository or model weights were released, and the work is a preprint accepted at KDD 2026 rather than a peer-reviewed journal article.

#Applications

GEARS is aimed at researchers who have rich scRNA-seq atlases but lack spatial context for the profiled cells, and who want to reconstruct plausible tissue geometry to study spatial organization, neighborhood composition, and putative cell-cell interactions. Because it does not require cell-type labels or histology, it is suited to exploratory analyses of new tissues, integration of legacy dissociated datasets with newer spatial references, and workflows where paired imaging is unavailable. Reconstructed coordinates can feed downstream spatial analyses such as neighborhood enrichment, spatial domain detection, and ligand-receptor inference that otherwise require natively spatial data.

#Impact

By reframing scRNA-seq-to-spatial mapping as direct geometry reconstruction rather than spot assignment or classification, GEARS contributes to a growing line of generative approaches for spatial single-cell inference and offers a label-free alternative to methods that depend on cell-type annotations or histological images. Its acceptance at KDD 2026 signals interest from the broader machine-learning community in spatial biology problems. As a recent preprint, its real-world adoption and robustness across diverse tissues and ST platforms remain to be established, and the absence of a released implementation at preprint time is a practical limitation for independent reproduction and use. The shared "GEARS" name with the well-known perturbation-prediction model may also create citation ambiguity that readers should be mindful of.