SQUINT: Spatial Quantization for Understanding and IN-painting Tissues

Generative modeling of spatially-resolved transcriptomics (SRT) at single-cell resolution is a key challenge in genomics towards profiling microenvironments in tissues with various biological and clinical applications. Typically, graph auto- encoders have been employed for learning continuous cell representations towards downstream tasks such as niche identification. However, these provide limited utility for modeling SRT data at the tissue level. To address this challenge, we propose a novel method called SQUINT which uses vector quantization on latent cell embeddings obtained from graph neural networks. By masking gene expression for a fraction of cells during training and conditioning on relative spatial distances, SQUINT learns a set of informative codes that can be used as cell tokens to model tissues as token sequences. During inference, SQUINT uses these codes to impute gene expression of unseen microenvironments at user-specified spatial locations in a tissue section and outperforms a benchmark method for generating gene expression. We further showcase the translational relevance of these codes capturing meaningful tissue structures beyond individual cells through downstream tasks such as 3D imputation, tumour stratification, and cell-type perturbation.