SPATIA: Multimodal Generation and Prediction of Spatial Cell Phenotypes

SPATIA is a multimodal deep learning framework from the Harvard MIMS lab that jointly models cell imagery and gene expression to predict and generate spatial cell phenotypes. It uses cross-attention to fuse morphology with transcriptomics and transformer modules to aggregate niche- and tissue-level context, and can generate cell morphology images conditioned on predicted state transitions via flow matching.

SPATIA (accepted at ICML'26) tackles a gap in spatial transcriptomics analysis: most pipelines treat cell imagery and gene expression as separate modalities. SPATIA fuses them via cross-attention, offering dual encoder variants (SPATIA-scprint, SPATIA-scgpt) and multi-scale transformer aggregation across niche and tissue levels.

Multimodal fusion of morphology + transcriptomics
Multi-scale niche/tissue transformer aggregation
Generative image synthesis conditioned on predicted cell-state transitions (flow matching)