BioLACE: unifying spatial geometry and marker priors for cohesive cell-type clustering in spatial transcriptomics

Abstract

Spatial transcriptomics (ST) provides high-dimensional gene expression profiles together with spatial coordinates, enabling the reconstruction of tissue architecture at cellular resolution. While recent graph-based deep learning methods have advanced spatial clustering in ST, many rely on complex architectures that obscure interpretability and rarely integrate biological priors such as marker gene information. We introduce BioLACE, a scalable framework that unifies spatial structure, transcriptomic variation, and curated marker gene profiles within a shared Variational Autoencoder (VAE) latent space. BioLACEjointly optimizes three complementary objectives: (1) a VAE reconstruction loss to preserve transcriptional structure, (2) a graph Laplacian regularizer to enforce spatial smoothness, and (3) a temperature-scaled InfoNCE contrastive loss guided by marker-informed similarities. Applied to MERFISH hypothalamus, mouse spinal cord, and Slide-seq mouse cerebellum datasets, BioLACE achieves superior cell type clustering accuracy, well-defined biologically consistent boundaries, and interpretable latent representations, highlighting its generality and scalability for modern ST analysis. The source code and tutorials for BioLACE are publicly available at https://github.com/maiziezhoulab/BioLACE.