GenAR: Next-scale autoregressive generation for spatial gene expression prediction

Area of Science:

• Computational biology
• Genomics
• Biomedical imaging

Background:

• Spatial Transcriptomics (ST) provides spatially resolved gene expression data but is expensive.
• Hematoxylin and Eosin (H&E) images are widely available and offer a cost-effective alternative for expression prediction.
• Existing computational methods often predict genes independently and use continuous regression, leading to biologically implausible results.

Purpose of the Study:

• To develop a cost-effective computational framework for predicting gene expression from H&E images.
• To address limitations of current methods by modeling gene co-expression and discrete expression counts.
• To improve the biological plausibility and utility of gene expression predictions for downstream analyses.

Main Methods:

• Introduced GenAR, a multi-scale autoregressive framework for gene expression prediction.
• Clustered genes hierarchically to capture cross-gene dependencies.
• Modeled gene expression as discrete count tokens and conditioned predictions on fused histological and spatial embeddings.

Main Results:

• GenAR achieved state-of-the-art performance across five diverse ST datasets.
• The framework successfully predicted gene expression from H&E images, modeling discrete counts.
• The coarse-to-fine approach ensured principled conditional decomposition and biologically plausible outputs.

Conclusions:

• GenAR offers a cost-effective and accurate method for predicting spatially resolved gene expression from H&E images.
• The discrete count modeling approach overcomes limitations of continuous regression, enhancing biological relevance.
• GenAR has significant implications for precision medicine and cost-effective molecular profiling.