HierarchicalDAEW: Domain-Aware Edge-Weighted Graph Convolution with Evidential Uncertainty for Multi-Section Spatial Gene Expression Prediction from H&E Histology

Predicting spatially resolved gene expression from standard H&E histology slides is a critical task for clinical pathology, as it offers a low-cost alternative to expensive spatial transcriptomics assays. However, existing methods often treat tissue spots as independent or fail to account for the structural heterogeneity of tissue. The authors introduce HierarchicalDAEW, a dual-graph architecture designed to explicitly encode tissue architecture and quantify prediction reliability.

The model operates on two levels. First, a spot-level graph uses a Domain-Aware Edge-Weighted (DAEW) convolutional operator. Unlike standard graph neural networks, this operator learns separate projections for intra-domain, inter-domain, and boundary edges, which are identified using Leiden clustering. This allows the model to adapt its message-passing mechanism to the specific biological context of the tissue. Second, a gene-level graph integrates protein-protein interaction priors with tissue-specific co-expression data to refine predictions across a broader gene panel. Finally, the model employs evidential uncertainty estimation to provide calibrated confidence intervals for every prediction, flagging low-confidence spots for expert review.

HierarchicalDAEW was evaluated against thirteen published baselines across six human tissue sections (breast, colorectal, prostate, and cerebellar). The model consistently achieved the highest correlation with ground-truth gene expression. Ablation studies confirmed that both the domain-aware edge typing and the hierarchical structure are essential for this performance. Furthermore, the evidential uncertainty estimation proved superior to Monte Carlo dropout, providing near-exact empirical coverage and enabling the model to function as a reliable clinical decision-support tool rather than a black-box classifier.

By predicting gene expression from ubiquitous H&E images, this method enables scalable molecular profiling without the high cost of specialized spatial transcriptomics assays. The inclusion of calibrated uncertainty quantification addresses a major barrier to clinical adoption, as it allows pathologists to distinguish between high-confidence predictions and those that require further verification, effectively bridging the gap between computational prediction and clinical decision-making.