HiFi-ST

HiFi-ST is a framework for predicting spatial gene expression from tissue histology by reconstructing a continuous expression field rather than a set of discrete per-spot values. It was introduced in the bioRxiv preprint "HiFi-ST: High-Fidelity Reconstruction of Continuous Spatial Transcriptomic Expression Fields via Conditional Neural Fields" (posted July 2026) by Hui Li, Lei Tang, Wei Han, Xiaodong Yang, and Xiaowei Chen at Nanjing Drum Tower Hospital, the teaching hospital affiliated with Nanjing University Medical School.

Most histology-to-expression methods regress a single expression vector for each sequenced spot, treating spots as isolated points on a grid. HiFi-ST instead models expression as a continuous function of spatial position, conditioned on the underlying tissue morphology. Each measured spot is treated as a regional observation covering an area of tissue rather than a point sample, which the model reconstructs by integrating the continuous field over that region. This formulation is designed to counter the spatial aliasing and information loss that arise when area measurements are collapsed to point predictions.

Positioned within the growing body of work that infers molecular readouts from routine imaging, HiFi-ST joins histology-to-expression models such as Phoenix and MoLF, but takes a distinct implicit-representation approach that can be queried at arbitrary spatial coordinates.

#Key Features

• Conditional neural field representation: Expression is represented as a continuous coordinate-based function conditioned on tissue image features, enabling prediction at spatial locations beyond the original measurement grid.
• Multiscale tissue feature extraction: Histology features are extracted at multiple scales so that both fine cellular detail and broader tissue context inform each prediction.
• FiLM conditioning: Feature-wise linear modulation injects the multiscale morphological features into the neural field, making the reconstructed field a function of the observed tissue rather than an unconditioned per-slide fit.
• Regional observation model: Each sequenced spot is modeled as an area-integrated observation, with Monte Carlo integration over the spot's support domain used to match the continuous field to measured expression.

#Technical Details

HiFi-ST couples a coordinate-based neural field with a multiscale histology encoder. Morphological features drive the field through FiLM conditioning, and per-spot predictions are formed by Monte Carlo integration of the continuous field across each spot's regional support domain, aligning the reconstruction with the area that spatial transcriptomics spots physically sample. The model is trained once and evaluated on three independent public paired histology and spatial transcriptomics datasets: HER2-positive breast cancer (HER2+), cutaneous squamous cell carcinoma (cSCC), and the Alex_NatGen cohort. It is compared head-to-head with feed-forward baselines including HisToGene, His2ST, BLEEP, THItoGene, and mclSTExp. The preprint is released under a CC BY license.

#Applications

By reconstructing a continuous expression field, HiFi-ST supports spatially-resolved molecular analysis of tissue from histology alone, including tumor-microenvironment characterization and the identification of candidate tertiary lymphoid structure (TLS) regions. Because the field can be queried between measured spots, it lends itself to higher-resolution readouts than the sequencing grid provides, of interest to researchers in computational pathology and spatial biology who want molecular context without commissioning additional spatial assays.

#Impact

HiFi-ST reframes histology-to-expression prediction as continuous-field reconstruction, arguing that treating spots as area observations rather than points better preserves spatial information. As a preprint, its results await peer review and independent validation. No public code repository or released model weights accompany the current posting, which limits independent reproduction and downstream reuse until such artifacts are made available.