Research & Papers

Path2ST predicts gene expression from H&E images with SOTA accuracy

Turns cheap tissue slides into spatial transcriptomics data, no special equipment needed.

Deep Dive

Spatial transcriptomics (ST) maps gene expression within tissue, but it's expensive and technically demanding. Path2ST, developed by Ruochen Liu and Wei Lou, sidesteps this by predicting gene expression directly from hematoxylin and eosin (H&E) stained images—the standard, low-cost tissue slides used in pathology. The paper, published on arXiv, treats H&E-to-ST as a cross-modal semantic translation task, capturing the biological hierarchy where cell types form tissue microenvironments that drive gene expression.

The framework introduces three innovations: a hierarchical cell-tissue conditioning mechanism that fuses cellular features with tissue-level semantics, a scale-adaptive autoregressive generation process that synthesizes expression in a coarse-to-fine manner, and SpectraLoss, a loss function that enforces ordinal fidelity and models transcriptional bursts. Tested on three datasets, Path2ST outperforms existing methods in accuracy and spatial coherence. The authors have released the code, enabling other labs to adopt it immediately for research.

Key Points
  • Path2ST predicts spatial gene expression from H&E images, eliminating the need for costly spatial transcriptomics experiments.
  • Uses hierarchical cell-tissue conditioning and scale-adaptive autoregressive generation for biologically consistent expression synthesis.
  • Achieves state-of-the-art results on three datasets; code available on GitHub.

Why It Matters

Path2ST could democratize spatial transcriptomics, letting any lab with standard pathology slides derive gene expression maps at near-zero marginal cost.

📬 Get the top 10 AI stories daily