SPARC: Multi-view Spatial Transcriptomics Clustering via Prototypical Contrast and Attentional Fusion

ICLR 2026 Conference Submission17684 Authors

Keywords: Prototypical Contrastive Learning, Multimodal Attention Mechanism, Adversarial Training, Graph Convolutional Network, Self-supervised Learning

Abstract: Spatial transcriptomics (ST) technologies measure gene expression along with spatial coordinates, enabling integrative analysis of tissue structure and function. However, existing approaches struggle to fully exploit multi-view complementarity and suffer from an objective mismatch between common pre-training tasks and downstream clustering. We propose SPARC (Spatial transcriptomics clustering via Prototypical Contrast and Attentional Fusion), a unified framework with a multi-level alignment strategy. SPARC introduces three key innovations: Feature-level Alignment via a multimodal attention mechanism that dynamically fuses gene expression and tissue morphology; Distribution-level Alignment via adversarial training to bridge modality gaps in the latent space; and Semantic-level Alignment via a prototype-driven contrastive objective that aligns pre-training with clustering by contrasting samples against learnable semantic prototypes rather than instances, mitigating class-collision. Coupled with a multi-branch GNN and multi-task reconstruction (ZINB for counts and image reconstruction), SPARC yields robust, cluster-friendly embeddings and significantly improves spatial domain identification.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 17684