BMCST: Balanced multi-view clustering for spatially resolved transcriptomics with Mamba-driven dynamic feature refinement

Abstract

Spatially resolved transcriptomics (SRT) provides histological images, spatial location, and gene expression profiles for spatial clustering analysis, offering profound insights into cellular interactions and disease progression mechanisms. Despite the progress made in spatial clustering research, several primary challenges persist due to the inherent noise and view heterogeneity in SRT data: (1) Intra-view: existing methods are prone to overlooking the interference from low-quality or redundant features when modeling global dependencies among spots, which leads to excessive information propagation and redundant computations; (2) Inter-view: the widely adopted joint training paradigm tends to result in an imbalance and suboptimal optimization of view-specific features. To this end, we propose a novel balanced multi-view clustering method for SRT data, referred to as BMCST. Specifically, we introduce a state-adaptive processing architecture, the Mamba-driven Dynamic Feature Refinement (MDFR) module, which adapts to the state of the input to dynamically select and prioritize the most informative features within the intra-view context, disregarding the noise and irrelevant information. This strategy ensures comprehensive global modeling while precisely capturing local spatial dependencies. Additionally, an unsupervised dominant view mining mechanism is introduced to dynamically identify the most discriminative perspectives prior to the feature fusion process, coupled with optimized alignment among views and consistent similarity distributions between nodes, aiming to mitigate inter-view information imbalance. Extensive experiments show that the proposed BMCST outperforms other state-of-the-art methods in spatial domain identification.