Trustworthy super-resolution reconstruction across spatial omics modalities

Abstract

Sequencing-based spatial omics platforms provide scalable and unbiased profiling of transcriptomic, epigenomic, and isoform-resolved signals, but their spatial resolution remains limited because each capture unit aggregates molecular information from multiple cells. Existing computational enhancement methods reconstruct high-resolution spatial gene expression from spot-level data by integrating histology, but most are designed primarily for spatial transcriptomics, rely predominantly on histological features, and lack appropriate validation strategies, leading to overfitting, spurious spatial patterns, and limited reliability. Here we present spEnhance, a generalizable and trustworthy computational framework for super-resolution enhancement of spot-level spatial omics data. spEnhance integrates histological features with complementary molecular information, including single-cell RNA-seq references and gene co-expression structure, to reconstruct high-resolution spatial molecular profiles. To enable reliable model selection from a single tissue section, spEnhance introduces a count-splitting strategy that generates statistically independent training and validation sets from spot-level measurements. spEnhance further quantifies prediction reliability through predictive residuals, providing an interpretable proxy for spatially resolved uncertainty. Comprehensive benchmarking across multiple tissues, platforms, and molecular modalities demonstrates that spEnhance achieves state-of-the-art accuracy, recovers fine-grained tissue structures, mitigates overfitting, and provides calibrated reliability estimates. Beyond spatial transcriptomics, spEnhance extends to isoform-level, epigenomics, proteomic and metabolomic spatial omics. Collectively, spEnhance establishes a modality-general and trustworthy framework for enhancing spatial omics data, enabling more accurate and reliable investigation of spatially resolved molecular regulation.