GeneFlow reads 1,000 genes straight from H&E slides, cutting inference to 0.31 GFLOPs
Synopsis
Predicting spatial gene expression from H&E histology has relied on complex architectures and multi-stage pipelines, usually limited to fixed panels of about 200 genes. GeneFlow pairs flow-matching generative modeling with histology and gene tokenizers to predict each spot's expression profile directly, reaching the highest PCC across four public datasets, a 17.1% gain on PRAD PCC-1000, with a 0.31 GFLOPs trainable module and 0.61 s sampling per sample.
Interpretation
Spatial gene expression prediction can now cover a 1,000-gene panel instead of a fixed subset of roughly 200 genes. Earlier methods relied on complex architectures and multi-stage pipelines, with gene panels often limited to 200 genes, while practical use typically needs coverage of 1,000 or more. On four public HEST-1k datasets, GeneFlow scores higher than six baselines, StNet, EGN, BLEEP, TRIPLEX, Stem and STFlow, on every PCC metric, with a 17.1% PCC-1000 gain over the best baseline on PRAD.
The model predicts each spot's expression directly from histology, without iterative denoising and without explicitly modeling spot-to-spot spatial relations. Diffusion-based methods need multi-step iterative sampling, and some approaches rely on multi-scale branches or explicit cell-cell or spot-spot relation modeling. The trainable prediction module costs 0.31 GFLOPs and 0.61 s per sample, below STFlow at 53.35 GFLOPs and Stem at 42.79 GFLOPs, and sampling stabilizes at 2 to 5 steps.
Histology features come from the pretrained pathology foundation models UNI and CONCH, while genes are compressed into a small set of gene module tokens. Earlier methods either trained image encoders from scratch or treated genes as largely independent regression targets. In ablation, removing the gene tokenizer drops PCC-1000 from 0.2874 to 0.1997, removing CONCH drops it to 0.0645, and removing UNI drops it to 0.0374.
Perspective
Evaluation is confined to four spot-based spatial transcriptomics datasets in HEST-1k, kidney, HER2ST, PRAD and mouse brain, each holding out one slide for testing and training on the rest, with the top 1,000 highly expressed and highly variable genes from training slides as targets.
Prediction accuracy is constrained by the quality and representation capacity of the pathology foundation model; the method relies mainly on paired H&E and spatial transcriptomics data for supervision and does not explicitly incorporate external biological knowledge, which may limit generalization where paired data are scarce. PCC gains saturate after roughly 7.5k training steps, with the best result at 15k steps; PCC jumps from 1 to 2 sampling steps and then stabilizes.
