Gene-DML: Dual-Pathway Multi-Level Discrimination for Gene Expression Prediction from Histopathology Images

TL;DR

Gene-DML addresses the challenge of predicting spatial gene expression from histopathology by introducing a dual-pathway framework that performs multi-scale instance-level alignment and cross-level instance-group alignment between WSIs and gene profiles. The method leverages local, neighbor, and global image scales with a shared encoder, supplemented by feature grouping and bidirectional instance-group guidance to enforce semantic consistency across modalities. Extensive CV and external Visium tests demonstrate state-of-the-art accuracy and strong generalization, with ablations confirming the complementary value of each module. This approach advances cross-modal learning in spatial transcriptomics, enabling more robust molecular profiling from morphological data and facilitating biomarker discovery.

Abstract

Accurately predicting gene expression from histopathology images offers a scalable and non-invasive approach to molecular profiling, with significant implications for precision medicine and computational pathology. However, existing methods often underutilize the cross-modal representation alignment between histopathology images and gene expression profiles across multiple representational levels, thereby limiting their prediction performance. To address this, we propose Gene-DML, a unified framework that structures latent space through Dual-pathway Multi-Level discrimination to enhance correspondence between morphological and transcriptional modalities. The multi-scale instance-level discrimination pathway aligns hierarchical histopathology representations extracted at local, neighbor, and global levels with gene expression profiles, capturing scale-aware morphological-transcriptional relationships. In parallel, the cross-level instance-group discrimination pathway enforces structural consistency between individual (image/gene) instances and modality-crossed (gene/image, respectively) groups, strengthening the alignment across modalities. By jointly modeling fine-grained and structural-level discrimination, Gene-DML is able to learn robust cross-modal representations, enhancing both predictive accuracy and generalization across diverse biological contexts. Extensive experiments on public spatial transcriptomics datasets demonstrate that Gene-DML achieves state-of-the-art performance in gene expression prediction. The code and processed datasets are available at https://github.com/YXSong000/Gene-DML.

Figures (8)

• Figure 1: An overview of Gene-DML framework. Pairs of WSI tiles and gene profiles are aligned from both the multi-scale instance-level and the cross-level instance-group discrimination. An illustration of feature grouping with instance-group alignment is shown in \ref{['fig2']}.
• Figure 2: Detailed illustration of feature grouping intuition. The image group features (gradient color squares) and gene group features (red squares) are grouped into $k$ groups with centroids $\{C^I_1, \dots C^I_k\}$ and $\{C^G_1, \dots C^G_k\}$.
• Figure 3: Visualization of cancer biomarker gene expression level prediction. All predicted values of gene expression are normalized to the range $[0,1]$. The values in parentheses are the PCC between the ground truth of gene expression and the prediction on genes GNAS and FASN.
• Figure 4: Visualization of cancer biomarker genes GNAS and FASN prediction on the dataset HER2ST.
• Figure 5: Visualization of cancer biomarker genes GNAS and FASN prediction on the dataset STNet.