From histology to spatial transcriptomics: establishing a lightweight single-patch baseline

Abstract

Abstract Predicting spatial gene expression from hematoxylin and eosin (H&E)-stained histological images is a central challenge in the field of computational pathology. Although many models leverage the spatial context from entire whole-slide images to boost performance, the predictive capability achievable from single, isolated tissue patches remains unexplored. Without this fundamental baseline, it is difficult to fairly measure the marginal benefit of incorporating spatial context. In this study, we established a lightweight, efficient, and reproducible baseline for single-patch gene expression prediction, thereby providing a standardized foundation for future research. We trained various convolution-based pretrained architectures, such as EfficientNet, ResNet, and DenseNet, using a fixed data split and seed on a curated human liver Visium dataset. To assess cross-tissue generalizability, we further validated the proposed baseline on an independent breast cancer spatial transcriptomics dataset. Through full fine-tuning and morphology-preserving augmentation, EfficientNet-B0 achieved a maximum Pearson correlation coefficient (PCC) of 0.310 for highly expressed genes on liver tissue, surpassing previous single-patch methods. Furthermore, the model predicted 50 genes with PCC $$\ge $$ 0.30 using only 5.3M parameters-−2.5 $$\times $$ more genes than a ResNet-50-based baseline despite using 5 $$\times $$ fewer parameters–suggesting a favorable performance-efficiency tradeoff within this single-patch setting. On the independent breast cancer dataset, EfficientNet-B0 consistently achieved best or near-best performance across all gene sets under both single-donor and multi-donor evaluations, demonstrating stable generalization across tissue types. Subsequent biological analysis reveals that model prediction accuracy is primarily driven by the strength of spatial organization in H&E histology, measured by Moran’s I spatial autocorrelation, rather than transcript abundance or chemical properties. Consequently, this work provides a reference for quantifying the contributions of spatial context, multiresolution integration, and architectural complexity in future spatial transcriptomics models.

Ähnliche Arbeiten