From features to slice: parameter-cloud modeling of spatial transcriptomics for simulation and 3D interpolatory augmentation.

Computational simulation and data augmentation of spatial transcriptomics (ST) are essential for quantitative benchmarking, reproducibility, and methodological innovation. Yet current models often lack flexibility in controlling spatial and transcriptional heterogeneity, fail to capture higher-order gene dependencies, and rarely extend to three-dimensional or alignment-aware contexts. Here we present FEAST, a computational infrastructure that models ST data within a parameter cloud - a latent manifold encoding gene-level mean, variance, and sparsity. By sampling and perturbing this manifold, FEAST generates high-fidelity synthetic slices with tunable spatial and transcriptional variation, enabling systematic evaluation of clustering, deconvolution, and spatial alignment algorithms. Beyond two dimensions, FEAST performs 3D parameter-cloud interpolation guided by optimal transport and benchmark alignment, reconstructing continuous tissue architectures while preserving molecular coherence. Together, these capabilities establish FEAST as a foundational platform for standardized benchmarking, data augmentation, and 3D reconstruction in spatial transcriptomics. The source code and tutorials for FEAST are publicly available at https://github.com/maiziezhoulab/FEAST, and can be installed via Pypi at https://pypi.org/project/FEAST-py/.