More Genes, Faster Inference: Efficient Large-Panel Spatial Transcriptomics Prediction from Histology
Abstract
Spatial transcriptomics (ST) prediction from histology aims to infer spatially resolved gene expression profiles directly from tissue images. This is a challenging problem that is currently often addressed with complicated model architectures, additional inputs, and complicated training procedures, while remaining limited to fixed, small gene panels. This raises a natural question: Is there an efficient method for large-panel ST prediction? Here, we present GeneFlow, a simpler yet effective framework for large-scale spatial gene expression prediction that unifies flow-matching generative modeling with structured multimodal fusion. Specifically, GeneFlow first encodes both histological morphology and cross-gene interactions in spatial spots through histology and gene tokenizers. The model directly predicts spatial gene expression, rather than an intermediate velocity field, while being trained with a velocity-based flow-matching objective. We show that, even without relying on heavy generative pipelines, GeneFlow achieves state-of-the-art performance across multiple ST benchmarks, with up to 56.8% gains in 1,000-gene prediction. Extensive experiments further show that GeneFlow produces biologically meaningful spot profiles while reducing computational cost by orders of magnitude, thereby improving efficiency and scalability, particularly in clinical settings. We believe GeneFlow provides an opportunity for leveraging widely available histology images for downstream biological discovery. Code is available at: https://github.com/ZZhangsm/GeneFlow
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
GeneFlow introduces an efficient flow-matching framework for large-panel spatial transcriptomics prediction from histology, achieving state-of-the-art accuracy with significantly reduced computational costs. The paper makes a strong contribution to the intersection of computer vision and computational biology by demonstrating that generative models can be streamlined to handle high-dimensional gene expression data effectively, offering a scalable solution for clinical and research applications where traditional ST methods are prohibitively expensive.
Methodology
The paper proposes GeneFlow, a framework for spatial transcriptomics (ST) prediction from histology images. The core methodological contribution is the application of flow-matching generative modeling to this specific domain. The authors claim a simplification by directly predicting spatial gene expression rather than an intermediate velocity field, while still utilizing a velocity-based flow-matching objective. The architecture involves separate tokenizers for histological morphology and cross-gene interactions. While the application of flow matching to ST is a logical extension of recent generative trends, the specific architectural choices (dual tokenizers) and the claim of "simpler" training are somewhat standard in the context of modern diffusion/flow models. The novelty lies primarily in the efficient integration for large-scale panels rather than a fundamental shift in the generative paradigm.
Results & evaluation
The paper reports state-of-the-art performance across multiple ST benchmarks, claiming up to 56.8% gains in 1,000-gene prediction. It also highlights significant reductions in computational cost, which is a critical practical advantage for clinical scalability. The evaluation appears to cover both accuracy and efficiency, which are the two main pain points in this subfield. However, without access to the full experimental tables (as the text provided is a skeleton/abstract-heavy version), the depth of ablation studies and comparison against recent baselines (like ST-Net, HistoGene, etc.) is assumed to be standard for a KDD paper. The magnitude of improvement (56.8%) is substantial and suggests a strong empirical result.
Reproducibility
The authors provide a GitHub repository link (https://github.com/ZZhangsm/GeneFlow), which is a positive indicator for reproducibility. The code availability allows for verification of the claimed efficiency gains and accuracy improvements.
Limitations
The primary limitation is the reliance on histology images, which may not capture all biological nuances present in raw transcriptomic data. Additionally, the "large-panel" prediction is still constrained by the quality of the histology-to-gene mapping, which can be noisy. The paper's claim of "simpler" architecture needs to be weighed against the complexity of tuning flow-matching hyperparameters.
Broader impact
This work has significant potential impact on computational pathology and genomics. By enabling efficient, large-panel ST prediction from widely available histology slides, it could democratize access to spatial gene expression data, which is currently expensive and labor-intensive to generate via wet-lab methods. This could accelerate drug discovery and cancer research by providing richer spatial context to existing histological archives. GeneFlow introduces an efficient flow-matching framework for large-panel spatial transcriptomics prediction from histology, achieving state-of-the-art accuracy with significantly reduced computational costs. The paper makes a strong contribution to the intersection of computer vision and computational biology by demonstrating that generative models can be streamlined to handle high-dimensional gene expression data effectively, offering a scalable solution for clinical and research applications where traditional ST methods are prohibitively expensive.
Author of this paper? Send a correction.