Impact of Data Quality on Deep Learning Prediction of Spatial Transcriptomics from Histology Images

Spatial transcriptomic technologies enable high-throughput quantification of gene expression at specific locations across tissue sections, facilitating insights into the spatial organization of biological processes. However, high costs associated with these technologies have motivated the development of deep learning methods to predict spatial gene expression from inexpensive hematoxylin and eosin-stained histology images. While most efforts have focused on modifying model architectures to boost predictive performance, the influence of training data quality remains largely unexplored. Here, we investigate how variation in molecular and image data quality stemming from differences in spatial transcriptomic technologies impact deep learning-based gene expression prediction from histology images. To identify the aspects of data quality that impact predictive performance, we conducted in silico ablation experiments, which showed that increased sparsity and noise in molecular data degraded predictive performance, while in silico rescue experiments via imputation provided only limited improvements that failed to generalize beyond the test set. Likewise, reduced image resolution can degrade predictive performance and further impacts model interpretability. We further demonstrate that these data quality-driven effects are reproducible across multiple spatial transcriptomics datasets and remain consistent when using alternative feature extractors and model architectures. Overall, our results show how improving data quality provides an orthogonal strategy to tuning model architecture in spatial transcriptomics-based predictive modeling, highlighting the need to account for technology-specific limitations that directly impact data quality when developing predictive methodologies.