Research map / Graph learning / Knowledge and temporal graphs
PRAGA: Prototype-aware Graph Adaptive Aggregation for Spatial Multi-modal Omics Analysis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PRAGA, a spatial multi-modal omics framework that learns dynamic omics-specific feature graphs combined with spatial KNN graphs and trains them with a Bayesian GMM-based dynamic prototype contrastive loss to integrate RNA, ADT and ATAC data without known cell-type counts.
Problem
Fixed KNN graphs in spatial multi-modal omics miss latent semantic relations disturbed by sequencing perturbations, and spot annotations and cluster numbers are usually unknown, hindering representation learning.
Approach
Replace the fixed KNN feature graph with a learnable, KNN-initialized adjacency per modality, combine it with the spatial graph via learnable weights, supervise it with cross-modal reconstruction and a homogeneity (EMA reference) loss, and use Bayesian GMM split/merge to adaptively set prototypes for contrastive learning.
Key results
- On the Human Lymph Node dataset, PRAGA reaches NMI 39.47% and ARI 28.28%, versus 36.07% and 23.83% for the second-best SpatialGlue.
- On the Spatial epigenome-transcriptome mouse brain dataset, PRAGA improves NMI to 39.37% from SpatialGlue's 37.83%, and ARI to 27.06% from 26.33%.
- Ablation on Human Lymph Node shows removing the reconstruction loss drops NMI from 39.47% to 35.15%, the largest drop among the ablated components.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PRAGA instead of SpatialGlue | Spatial multi-modal omics aggregation · Human Lymph Node | NMI: 36.07 -> 39.47 | Helped |
| Use PRAGA instead of SpatialGlue | Spatial multi-modal omics aggregation · Human Lymph Node | F1-Score: 38.79 -> 42.33 | Helped |
| Use PRAGA instead of SpatialGlue | Spatial multi-modal omics aggregation · Mouse Brain | NMI: 37.83 -> 39.37 | Helped |
| Use PRAGA instead of SpatialGlue | Spatial multi-modal omics aggregation · Mouse Brain | F1-Score: 33.01 -> 35.02 | Helped |
| Use PRAGA instead of SpatialGlue | Spatial multi-modal omics aggregation · Spatial Multi-modal Omics Simulation Dataset | NMI: 96.98 -> 99.06 | Helped |
| Use PRAGA instead of SpatialGlue | Spatial multi-modal omics aggregation · Spatial Multi-modal Omics Simulation Dataset | F1-Score: 98.21 -> 99.47 | Helped |
| Replace the learnable omics-specific graph with a KNN graph (ablation) | PRAGA · Human Lymph Node | NMI: 39.47 -> 38.69 | Hurt |
| Add homogeneity loss (ablation) | PRAGA · Human Lymph Node | NMI: 39.25 -> 39.47 | Helped |
| Add reconstruction loss (ablation) | PRAGA · Human Lymph Node | NMI: 35.15 -> 39.47 | Helped |
| Add dynamic prototype contrastive loss (ablation) | PRAGA · Human Lymph Node | NMI: 38.46 -> 39.47 | Helped |
Surprising findings
- Table of simulation results shows STAGATE and several multi-omics baselines near-zero AMI, while PRAGA and SpatialGlue reach about 97-99 on the simulation dataset.
Limitations
- Evaluated mainly on clustering-style metrics on a few spatial multi-omics datasets; no downstream biological validation reported in main text.
- Relies on KNN initialization and a Bayesian GMM split/merge procedure whose hyperparameters (alpha, tau, beta) require tuning.
Experimental setup
Datasets: Human Lymph Node dataset Spatial epigenome-transcriptome mouse brain dataset Mouse thymus stereo-CITE-seq dataset SPOTS mouse spleen dataset Spatial multi-modal omics simulation dataset
Benchmarks: Human Lymph Node dataset Spatial epigenome-transcriptome mouse brain dataset Spatial multi-modal omics simulation dataset SPOTS mouse spleen (qualitative) Mouse thymus stereo-CITE-seq (qualitative)
Metrics: MI NMI AMI FMI ARI V-Measure F1-Score Jaccard Completeness
Baselines: MOFA+ CiteFuse TotalVI MultiVI STAGATE PAST SpatialGlue
Builds on: STAGATE SpatialGlue GCN
Closest papers by meaning
- Learning to Cluster Rare Cell Types: Implicit Semantic Data Augmentation for Spatial Multi-modal Omics Analysis (AAAI 2026 · similarity 0.89)
- GROVER: Graph-guided Representation of Omics and Vision with Expert Regulation for Adaptive Spatial Multi-omics Fusion (AAAI 2026 · similarity 0.89)
- Robust Integrative Analysis of Multi-omics Datasets via Nuclear-norm Maximization (AAAI 2026 · similarity 0.88)
- OgBench: A Framework for Evaluating Graph Neural Networks on Omics Data (NeurIPS 2026 · similarity 0.86)
- Multi-modal Topology-embedded Graph Learning for Spatially Resolved Genes Prediction from Pathology Images with Prior Gene Similarity Information (CVPR 2025 · similarity 0.85)
- Global Context-aware Representation Learning for Spatially Resolved Transcriptomics (ICML 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.