Predicting Gene Expression in Spatially Resolved Transcriptomics Across Samples Through Probabilistic Fusion of Hierarchical Histology and Spatial Information

Yinbo Liu, QiWu, Keyang Ye, Xiao He, Tian Tian

Keywords: Gene Expression Prediction, Cross-Slice Generalization, Variational Autoencoder, Spatially Resolved Transcriptomics

TL;DR: STevs: a deep generative model predicting gene expression from histology images through probabilistic fusion of hierarchical visual and spatial features, significantly improving cross-slice generalization and high-dimensional prediction

Abstract: Spatially resolved transcriptomics (SRT) is a transformative technology in biomedical research, yet its scalability is hindered by high costs and restricted capture areas. Computational methods for predicting high-quality gene expression are needed. However, existing methods are ineffective at predicting high-dimensional gene expression and generalizing to multiple spatial slices, primarily due to inter-sample heterogeneity and ineffective integration of visual and spatial information. To address these challenges, we propose STevs, a deep generative model designed to predict gene expression from tissue histology through a probabilistic fusion of image and spatial representations. STevs employs a multimodal variational autoencoder (VAE) architecture featuring parallel encoders that process distinct modalities: a Swin Transformer for hierarchical visual representation extraction and a multilayer perceptron (MLP) for spatial coordinates. The latent representations from these modalities are fused under uncertainty using a Product of Experts (PoE) mechanism. Furthermore, we introduce a latent alignment loss to explicitly promote a shared representation across modalities, thereby ensuring consistency between the image and spatial latent spaces. Comprehensive experimental evaluations demonstrate that STevs not only achieves state-of-the-art performance on standard within-slice gene prediction tasks but also significantly outperforms existing methods in the more challenging cross-slice prediction scenario. Our work provides a powerful computational tool capable of predicting gene expression directly from histology images, reducing the need for costly SRT experiments.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 24046