[Paper Review] PRAGA: Prototype-aware Graph Adaptive Aggregation for Spatial Multi-modal Omics Analysis

Xinlei Huang, Zhiqi Ma+6 more|arXiv (Cornell University)|Sep 19, 2024

Metabolomics and Mass Spectrometry StudiesBiochemistry, Genetics and Molecular Biology3 citations

TL;DR

PRAGA introduces a prototype-aware graph adaptive aggregation framework for spatial multi-modal omics analysis that learns dynamic, omics-specific graphs to capture latent semantic relations obscured by sequencing perturbations. By integrating Bayesian Gaussian Mixture Models for dynamic prototype contrastive learning, PRAGA achieves superior clustering performance, with improvements of up to 3.54% in F1-Score and 3.40% in NMI on real datasets, even without prior knowledge of cell type counts.

ABSTRACT

Spatial multi-modal omics technology, highlighted by Nature Methods as an advanced biological technique in 2023, plays a critical role in resolving biological regulatory processes with spatial context. Recently, graph neural networks based on K-nearest neighbor (KNN) graphs have gained prominence in spatial multi-modal omics methods due to their ability to model semantic relations between sequencing spots. However, the fixed KNN graph fails to capture the latent semantic relations hidden by the inevitable data perturbations during the biological sequencing process, resulting in the loss of semantic information. In addition, the common lack of spot annotation and class number priors in practice further hinders the optimization of spatial multi-modal omics models. Here, we propose a novel spatial multi-modal omics resolved framework, termed PRototype-Aware Graph Adaptative Aggregation for Spatial Multi-modal Omics Analysis (PRAGA). PRAGA constructs a dynamic graph to capture latent semantic relations and comprehensively integrate spatial information and feature semantics. The learnable graph structure can also denoise perturbations by learning cross-modal knowledge. Moreover, a dynamic prototype contrastive learning is proposed based on the dynamic adaptability of Bayesian Gaussian Mixture Models to optimize the multi-modal omics representations for unknown biological priors. Quantitative and qualitative experiments on simulated and real datasets with 7 competing methods demonstrate the superior performance of PRAGA. Code is available at https://github.com/Xubin-s-Lab/PRAGA.

Motivation & Objective

• Address the limitation of fixed KNN graphs in capturing latent semantic relations due to biological sequencing perturbations in spatial multi-modal omics data.
• Overcome the challenge of unknown spot annotations and cell type priors in real-world spatial omics applications.
• Develop a learnable, adaptive graph structure that integrates spatial proximity and cross-modal feature semantics to improve representation learning.
• Introduce a dynamic prototype contrastive learning mechanism based on Bayesian Gaussian Mixture Models to enable self-supervised representation optimization without requiring predefined cluster numbers.
• Achieve state-of-the-art performance in clustering and representation learning across simulated and real spatial multi-modal omics datasets.

Proposed method

• Construct a learnable, omics-specific graph that dynamically adapts to latent semantic relations between sequencing spots, replacing static KNN graphs.
• Integrate spatial coordinates and multi-modal omics features (e.g., transcriptomics, proteomics) into a unified graph-based representation learning framework.
• Employ a Bayesian Gaussian Mixture Model (BGMM) to model unknown cell type distributions and enable dynamic prototype learning in an open-set environment.
• Design a dynamic prototype contrastive learning loss ($\mathcal{L}_{dpcl}$) that adaptively identifies and optimizes cluster prototypes based on evolving feature representations.
• Incorporate homogeneity loss ($\mathcal{L}_h$) and reconstruction loss ($\mathcal{L}_{recon}$) to enhance feature consistency and representation fidelity.
• Use exponential moving average (EMA) with adaptive temperature scaling to stabilize the contrastive learning process and improve robustness.

Experimental results

Research questions

• RQ1 Can a learnable, dynamic graph structure outperform fixed KNN graphs in capturing latent semantic relations in spatial multi-modal omics data under biological noise?
• RQ2 How does dynamic prototype contrastive learning based on Bayesian Gaussian Mixture Models improve representation learning when cell type priors are unknown?
• RQ3 To what extent does PRAGA improve clustering performance across diverse spatial omics datasets compared to state-of-the-art methods?
• RQ4 How sensitive is PRAGA’s performance to hyperparameters such as initial cluster count, temperature, and loss weighting?
• RQ5 Can PRAGA maintain high performance even when the number of initial clusters deviates from the true number of cell types?

Key findings

• PRAGA achieves a 3.54% improvement in F1-Score and 3.40% in NMI on the Human Lymph Node dataset compared to baseline methods.
• On the Spatial epigenome–transcriptome mouse brain dataset, PRAGA improves F1-Score by 2.01% and NMI by 1.54%.
• In the simulation dataset, PRAGA achieves a 1.26% increase in F1-Score and a 2.08% gain in NMI.
• Abllation studies confirm that removing the learnable graph, homogeneity loss, reconstruction loss, or dynamic prototype contrastive loss leads to significant performance degradation.
• Parameter sensitivity analysis shows that PRAGA is robust to variations in hyperparameters, including initial cluster count, EMA speed ($\alpha$), temperature ($\tau$), and loss weight ($\beta$).
• When the initial cluster count matches the ground truth, PRAGA achieves the best performance; however, it remains effective even when the initial count is incorrect, demonstrating strong generalization under real-world uncertainty.