G2T: Tissue Reconstruction from Gene Expression via Embedding-Distance Flow Matching

Single-cell RNA sequencing, or scRNA-seq, offers high-resolution transcriptomic profiles of cells but disregards their spatial arrangement within tissues, which is crucial for understanding cell interactions and tissue organization. Spatial transcriptomics, or ST, provides spatial information but often compromises on gene panel coverage, resolution, or cost.

Researchers have developed a novel approach called G2T (Gene-to-Tissue), a generative deep learning model that reconstructs tissue from gene expression alone by predicting the pairwise distances between cells in an embedding space. G2T employs an attention-based Transformer with an output head that generates an Euclidean-Distance-Matrix (EDM), and it is trained using conditional flow matching.

This technique enables the network to denoise corrupt cell positions, conditioned on gene expression, by forecasting cell embeddings whose squared pairwise distances align with the true distance matrix. During inference, a fast locally-optimal-block (LOBPCG) multidimensional scaling step converts the predicted distance matrix into 2-D coordinates.

In a benchmark using a published MERFISH mouse primary motor cortex dataset, G2T outperforms the previous state-of-the-art method, LUNA, on all three standard metrics: Spearman correlation of pairwise-distance ranks, Contact F1, and per-cell-class Sum RSSD. G2T also demonstrates even more significant gains on a mouse central-nervous-system scRNA-seq atlas, when compared to an imputed spatial reference (STARmap PLUS-integrated locations, not the original coordinates).

By predicting cell geometry in a higher-dimensional embedding space instead of regressing 2-D coordinates, G2T overcomes the restrictive output parameterization of prior diffusion-based methods, creating a compact and scalable building block for tissue reconstruction from dissociated cells. This enables downstream analysis of spatial niches and cell-cell communication.