Multiomics Tissue Segmentation via Spatially-Informed Nested Biclustering Methods

2025-09-02

AI Summary

High-dimensional, spatially structured multi-omics mass spectrometry imaging (MSI) data—spanning lipids, peptides, and N-glycans—pose challenges for simultaneous tissue pixel segmentation and within-modality molecular signal clustering. Method: We propose Poseidon, a Bayesian nonparametric nested biclustering model that jointly integrates nested biclustering with a spatial hidden Markov random field (sHMMRF) to explicitly encode spatial dependencies among pixels, enabling multimodal synergistic analysis. To ensure scalability, we develop an efficient mean-field variational inference algorithm. Contribution/Results: Poseidon achieves high accuracy on synthetic benchmarks and, when applied to real MSI data from clear cell renal cell carcinoma, successfully identifies histopathologically relevant tissue regions and candidate biomarkers. It further uncovers hierarchical spatial–molecular association patterns across molecular modalities, providing mechanistic insights into tumor heterogeneity and microenvironment organization.

Technology Category

• Machine Learning: Multimodal Learning
• Computer Vision: Multi-modal Vision
• Data Mining & Knowledge Management: Mining of Visual, Multimedia & Multimodal Data

Application Category

• Graph Algorithms and Modeling for the Web: Algorithms and analysis for heterogeneous, signed, attributed, multi-relational, temporal, higher-order, and annotated Web-related graphs
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• User Modeling, Personalization and Recommendation: ML for personalized search and recommendations

Abstract

Matrix-Assisted Laser Desorption/Ionisation Mass Spectrometry Imaging (MSI) is a powerful technique for spatially resolved molecular profiling and cancer biomarker discovery. Recent advances, including a novel multiomics workflow, enable multiple rounds of MSI on the same tissue section, extracting diverse molecular classes, e.g., lipids, peptides, and N-glycans, while preserving spatial resolution. This innovation is particularly valuable for studies with limited tissue sample availability, such as rare diseases or small tumors. However, the resulting data are high-dimensional, spatially structured, and the various molecular types share the same pixel grid. To address these challenges, we propose Poseidon, a Bayesian nonparametric nested biclustering model that simultaneously segments the common tissue pixels and clusters molecular signals within each molecular class, leveraging the shared spatial structure. A separately exchangeable framework is first considered, and then extended to handle spatial data via hidden Markov random fields. For scalability, we implement an efficient mean-field variational inference algorithm tailored for multi-dataset analysis. After validating the efficacy of our method on simulated scenarios, we demonstrate the applicability of our model in a real-world case study, where multiomics measurements were performed on kidney tissue affected by clear cell renal cell carcinoma. The nested, hierarchical structure of Poseidon, combined with its principled inferential framework, allows the extraction of interesting biological insights, such as clear tissue segmentation and biomarker detection.

Problem

• Segments tissue pixels and clusters molecular signals simultaneously
• Handles high-dimensional spatially structured multiomics MSI data
• Addresses limited tissue sample availability in cancer biomarker discovery

Innovation

• Bayesian nonparametric nested biclustering model
• Hidden Markov random fields spatial extension
• Mean-field variational inference algorithm