HistoPrism: Unlocking Functional Pathway Analysis from Pan-Cancer Histology via Gene Expression Prediction

Predicting spatial gene expression from H&E histology offers a scalable and clinically accessible alternative to sequencing, but realizing clinical impact requires models that generalize across cancer types and capture biologically coherent signals. Prior work is often limited to per-cancer settings and variance-based evaluation, leaving functional relevance underexplored. We introduce HistoPrism, an efficient transformer-based architecture for pan-cancer prediction of gene expression from histology. To evaluate biological meaning, we introduce a pathway-level benchmark, shifting assessment from isolated gene-level variance to coherent functional pathways. HistoPrism not only surpasses prior state-of-the-art models on highly variable genes , but also more importantly, achieves substantial gains on pathway-level prediction, demonstrating its ability to recover biologically coherent transcriptomic patterns. With strong pan-cancer generalization and improved efficiency, HistoPrism establishes a new standard for clinically relevant transcriptomic modeling from routinely available histology.

💡 Research Summary

The paper introduces HistoPrism, a transformer‑based model designed to predict spatial gene expression directly from routine H&E whole‑slide images across multiple cancer types. Existing approaches either focus on single‑cancer settings, rely on complex multi‑stage pipelines, or evaluate performance solely on highly variable genes (HVGs) using Pearson correlation, which neglects functional coherence. HistoPrism addresses these gaps through three key innovations. First, it incorporates a cross‑attention module that injects a one‑hot cancer‑type embedding into the patch‑level visual features extracted by a pre‑trained pathology foundation model (PFM). This “pan‑cancer conditioning” enables the network to learn both shared and cancer‑specific morphological patterns. Second, the conditioned patch embeddings are processed by a lightweight transformer encoder (two layers, four attention heads) to capture long‑range spatial dependencies such as tumor boundaries and immune infiltration. Third, a simple multilayer perceptron regresses each transformer output to a vector of gene expression values, and the model is trained end‑to‑end with mean‑squared error loss. Compared with large BERT‑style foundations like STPath, HistoPrism uses far fewer parameters (≈12 M vs. >150 M) and requires substantially less GPU memory, making it more practical for clinical deployment.

Beyond model architecture, the authors propose a new evaluation framework called Gene Pathway Coherence (GPC). While traditional benchmarks assess correlation on the top‑N HVGs, GPC measures how well predicted expression recapitulates coordinated activity of biologically meaningful pathways. They curate 50 Hallmark gene sets and 87 Gene Ontology (GO) sets, filter for pathways containing 50–100 genes, and remove redundancy using a Jaccard similarity threshold of 0.1. For each pathway, the Pearson correlation between predicted and true expression is computed across all patches and averaged across slides, yielding a pathway‑level coherence score. This metric aligns model assessment with the scientific goal of understanding cellular function rather than merely fitting high‑variance signals.

Experiments are conducted on the large‑scale HEST1k dataset, which aggregates 153 cohorts from 36 studies, preserving inter‑center variability in staining, scanning, and spatial transcriptomics platforms. The dataset is split into training, validation, and hold‑out test sets stratified by cancer type. HistoPrism is compared against several baselines: the pan‑cancer foundation model STPath (using the GigaPath PFM), MLP with the UNI PFM, BLEEP, TRIPLEX, and two recent generative approaches (STEM diffusion and STFlow flow‑based). For fairness, the generative models are limited to the top 50 HVGs. Results show that HistoPrism outperforms all baselines on both traditional HVG Pearson correlation and the newly introduced GPC scores. Notably, on the GPC benchmark, HistoPrism achieves average pathway coherence of 0.71 for Hallmark sets (vs. 0.55 for STPath) and similarly improved scores for GO pathways, indicating a superior ability to recover biologically coherent transcriptional programs. In contrast, STEM and STFlow perform poorly in the pan‑cancer setting, highlighting current generative models’ difficulty in handling heterogeneous tumor biology.

The authors discuss limitations: the model currently predicts a subset of genes (≈5–10 k) rather than the full transcriptome, and rare gene prediction may need further refinement. The GPC benchmark itself depends on pathway size and redundancy thresholds, suggesting future work on standardizing functional evaluation. Nonetheless, HistoPrism’s combination of efficient architecture, cancer‑type conditioning, and pathway‑level validation establishes a new standard for clinically relevant transcriptomic inference from histology. The authors envision extensions to full‑transcriptome scaling, multimodal integration with clinical data, and real‑time deployment in pathology workflows to enable non‑invasive molecular profiling and personalized treatment decision‑making.