PEaRL: Pathway-Enhanced Representation Learning for Gene and Pathway Expression Prediction from Histology

TL;DR

PEaRL addresses the challenge of linking tissue morphology with molecular function by representing spatial transcriptomics through pathway activation scores computed with ssGSEA and aligning these pathway representations to histology via a transformer-based encoder and contrastive learning. The two-stage training workflow first learns a shared latent space for pathways and images, then trains lightweight heads to predict both pathway and gene expression, achieving superior PCCs across three cancer ST datasets and improving survival prognostics. Ablation studies highlight the importance of grounding transcriptomic signals in pathways and the advantage of using the UNI foundation model for histology features. Overall, PEaRL provides a more biologically faithful and interpretable multimodal framework that advances computational pathology beyond gene-level embeddings and paves the way for pan-cancer analyses and clinical biomarker discovery.

Abstract

Integrating histopathology with spatial transcriptomics (ST) provides a powerful opportunity to link tissue morphology with molecular function. Yet most existing multimodal approaches rely on a small set of highly variable genes, which limits predictive scope and overlooks the coordinated biological programs that shape tissue phenotypes. We present PEaRL (Pathway Enhanced Representation Learning), a multimodal framework that represents transcriptomics through pathway activation scores computed with ssGSEA. By encoding biologically coherent pathway signals with a transformer and aligning them with histology features via contrastive learning, PEaRL reduces dimensionality, improves interpretability, and strengthens cross-modal correspondence. Across three cancer ST datasets (breast, skin, and lymph node), PEaRL consistently outperforms SOTA methods, yielding higher accuracy for both gene- and pathway-level expression prediction (up to 58.9 percent and 20.4 percent increase in Pearson correlation coefficient compared to SOTA). These results demonstrate that grounding transcriptomic representation in pathways produces more biologically faithful and interpretable multimodal models, advancing computational pathology beyond gene-level embeddings.

Figures (10)

• Figure 1: Multimodal Learning with PEaRL. PEaRL considers gene-gene interactions via pathways which not only addresses the curse of high gene dimensionality but also acts as a molecular driver of tissue morphology leading to biologically meaningful representation learning between images and molecular data. We employ PEaRL for both gene and pathway expression prediction and also show its utility in survival analysis.
• Figure 2: Overview of PEaRL framework. The framework begins with extracting patches corresponding to spatial transcriptomics spots and feeding them to the image encoder to get image embeddings. We compute the pathway scores from the high dimensional gene expression data corresponding to the spatial transcriptomics spots and feed it as an input to the transformer based pathway encoder along with the positional embeddings. The image embeddings and pathway embeddings are aligned via contrastive learning.The learned image embeddings are then used for downstream gene and pathway expression prediction.
• Figure 3: Visualization of expression predictions of different pathways and their corresponding genes across three datasets compared with three baseline methods. We show Hallmark_allograft_rejection pathway and its gene HLA-DMB for the breast cancer case; Hallmark_epithelial_mesenchymal transition pathway and its gene QSOX1 for the skin cancer case; Reactome_ABC_family_proteins_mediated_transport and its gene DERL3 for the lymph carcinoma case.
• Figure 4: Visualization of the Leiden clusterings for ground truth (GT) and predicted gene expressions (for a skin cancer sample) using PEaRL and other baselines. ARI index shown in (.).
• Figure 5: Visualization of the pathway-pathway correlation plots across the three datasets and comparison of PEaRL with baseline models