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

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Current spatial transcriptomics (ST) simulation models lack flexibility in controlling spatial and transcriptional heterogeneity.
• Existing models fail to capture higher-order gene dependencies and rarely extend to 3D or alignment-aware contexts.
• There is a need for robust computational tools for quantitative benchmarking and reproducibility in ST.

Purpose of the Study:

• To present FEAST, a flexible computational infrastructure for modeling and generating synthetic spatial transcriptomics data.
• To enable systematic evaluation and benchmarking of ST algorithms through high-fidelity data augmentation.
• To extend ST data simulation and reconstruction capabilities to three-dimensional contexts.

Main Methods:

• FEAST models ST data using a parameter cloud, a latent manifold encoding gene-level mean, variance, and sparsity.
• It generates synthetic data by sampling and perturbing this manifold, allowing tunable spatial and transcriptional variation.
• FEAST employs 3D parameter-cloud interpolation guided by optimal transport for reconstructing continuous tissue architectures.

Main Results:

• FEAST generates high-fidelity synthetic ST slices with controllable heterogeneity.
• The tool enables systematic evaluation of clustering, deconvolution, and spatial alignment algorithms.
• FEAST successfully performs 3D reconstruction of tissue architectures while preserving molecular coherence.

Conclusions:

• FEAST provides a foundational platform for standardized benchmarking and data augmentation in spatial transcriptomics.
• The infrastructure supports methodological innovation by enabling flexible simulation of ST data.
• FEAST extends ST analysis to 3D, facilitating reconstruction of complex tissue architectures.