SPATIA

Multimodal Model for Prediction and Generation of Spatial Cell Phenotypes

---

Understanding how cellular morphology, gene expression, and spatial organization jointly shape tissue function is a central challenge in biology. SPATIA is a multi-scale model for spatial transcriptomics that:

• Learns cell-level embeddings by fusing image-derived morphological tokens and transcriptomic tokens via cross-attention

• Aggregates embeddings at niche and tissue levels with transformer modules to capture spatial context

• Generates cell morphology images conditioned on predicted state transitions using flow matching

---

Pipeline Overview

SPATIA is organized as a three-stage pipeline:

Stage	Component	Purpose
Stage 1	gene_encoders/	Representation learning (scPRINT or scGPT backbone)
Stage 2	generative_tasks/data_pairing_for_FM/	OT-based perturbation pair construction
Stage 3	generative_tasks/spatia_flow/	Flow-matching cell image generation

Downstream evaluation (clustering, annotation, biomarker prediction) lives in prediction_tasks/.

---

Getting Started

• Installation
  • Conda Environment
  • Package Installation
  • Generation Dependencies
  • Notes

Data & Training

• MIST Dataset Construction
  • Stage A: Crop Cell Images into LMDB
  • Stage B: Build Annotated h5ad and Register in lamindb
  • Stage C: Merge Per-Dataset LMDBs
  • Data Loading at Training Time
• Representation Training (Stage 1)
  • Prerequisites
  • SPATIA-scprint (Primary)
  • SPATIA-scgpt (Alternative)
  • Architecture Summary

Downstream Tasks

• Prediction Tasks
  • Table 2: Cross-Platform Clustering (Xenium + CosMx)
  • Table 3: Biomarker Prediction (HEST Benchmark)
  • Table 4: scRNA-seq Clustering & Annotation (GSE155468)
  • Labels and Reproducibility
• Generation Pipeline (Stages 2 & 3)
  • Stage 2: OT-Based Perturbation Pairing
  • Stage 3: Flow-Matching Image Generation
  • Niche-Level Generation

Reference

• Citation & Contact