HECLIP: histology-enhanced contrastive learning for imputation of transcriptomics profiles

Abstract

Motivation: Histopathology, particularly hematoxylin and eosin (H&E) staining, is pivotal for diagnosing and characterizing pathological conditions by visualizing tissue morphology. However, H&E-stained images inherently lack molecular resolution, necessitating costly and labor-intensive technologies like spatial transcriptomics (ST) to uncover spatial gene expression patterns. There is a critical need for scalable computational methods that can bridge this imaging-transcriptomics gap. Results: We present histology-enhanced contrastive learning for imputation of profiles (HECLIP), an innovative deep learning framework designed to infer spatial gene expression profiles directly from H&E-stained histology images. HECLIP employs an image-centric contrastive learning strategy to capture morphological features relevant to molecular expression. By minimizing dependence on ST data, HECLIP enables accurate and biologically meaningful predictions of gene expression. Extensive benchmarking on publicly available datasets demonstrates that HECLIP outperforms existing methods. Ablation studies confirm the contribution of each model component to its overall performance. Availability and implementation: The source code for HECLIP is freely available at: https://github.com/QSong-github/HECLIP.