Predicting Spatial Transcriptomics from Histology Images via Biologically Informed Flow Matching
Keywords: Spatial Transcriptomics; Histology Images
Abstract: Spatial transcriptomics (ST) has emerged as a promising technology to bridge the gap between histology imaging and gene expression profiling. However, its application to medical diagnosis is limited due to its low throughput and the need for specialized experimental facilities. To address this issue, we develop STFlow, a flow-based generative model to predict spatial transcriptomics from whole-slide histology images. STFlow is trained with a biologically-informed flow matching algorithm that iteratively refines predicted gene expression values, where we choose zero-inflated negative binomial distribution as a prior distribution to incorporate the inductive bias of gene expression data. Compared to previous methods that predict the gene expression of each spot independently, STFlow models the interaction of genes across different spots to account for potential gene regulatory effects. On a recently curated HEST-1k benchmark, we demonstrate STFlow substantially outperforms all baselines including pathology foundation models, with over 18% relative improvement over current state-of-the-art.
Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 11482
Loading