bio.rodeo
ModelsOrganizationsProvidersLeaderboardAboutSign in
bio.rodeo

The authoritative source for evaluating biological foundation models. No hype, just honest analysis.

Categories
  • DNA & Gene
  • RNA
  • Protein
  • Small molecule
  • Single-cell
  • Spatial omics
  • Pathology
  • Imaging
  • Metabolomics
  • Biosignals
  • Language model
bio.rodeoModelsOrganizationsProvidersLeaderboardAboutFAQSubmit a modelContact
© 2026 Pulsatance. All rights reserved. ~
Built by Pulsatance
models / pathology / stpath
PathologySpatial omics
Northeastern UniversityYale UniversityBroad InstituteReleased April 2025

STPath

Spatial transcriptomics prediction from H&E whole-slide images. One generative checkpoint covers 38,984 genes and 17 organs without fine-tuning.

28Openness

Where to run it

No providers recorded yet. Browse all providers

STPathPathologyNortheastern University

Spatial transcriptomics resolves gene expression while preserving each spot's position in a tissue section, but the sequencing itself remains low-throughput and expensive, so most archived pathology slides have no molecular readout at all. A natural workaround is to infer expression directly from the hematoxylin-and-eosin image, and several methods do exactly that by attaching a regression head to a pathology foundation model such as UNI or Prov-GigaPath. Those models are trained one dataset at a time, on one organ, one sequencing platform, and one gene panel — every new deployment requires retraining to match the target gene set.

STPath, from Northeastern University with Yale University and the Broad Institute, removes that constraint by pretraining a single model on a large paired corpus of whole-slide images and spatial transcriptomics profiles. The corpus assembles HEST-1K and STImage-1K4M into 983 slides covering 38,984 genes, 17 organs, and 4 sequencing technologies, with downstream evaluation samples and source-overlapping duplicates removed. The result is one checkpoint that predicts spatially resolved expression for arbitrary organs and gene sets without downstream fine-tuning.

The design difference that makes this work is generative rather than discriminative. Instead of learning a direct image-to-expression map, STPath is trained with a masked gene expression objective, recovering the profiles of hidden spots from the visible ones. That framing lets the model exploit gene-gene and spot-spot dependencies, and it enables in-context prompting at inference: supplying measured expression for a small fraction of spots sharpens predictions for the rest.

#Key Features

  • Organ- and platform-agnostic inference: A single pretrained checkpoint covers 17 organs and 4 sequencing technologies, with organ type and platform supplied as optional conditioning tokens that can be left unspecified.
  • Geometry-aware transformer: Each spot is a token whose representation fuses Prov-GigaPath image features, gene expression, organ type, and technology; a frame-averaging transformation biases the attention map by spot coordinates so the model is invariant to slide orientation.
  • In-context prompting: Because prediction is framed as masked-token recovery, measuring a few spots improves the rest. On the CCRCC dataset, prompting just 5% of spots raises Pearson correlation by 0.257.
  • Predicted expression as a slide-level feature: Concatenating STPath's inferred biomarker expression with visual features improves weakly supervised survival prediction by 6.1% for UNI and 5.6% for Prov-GigaPath, and mutation prediction by 5.6% and 5.0% respectively.

#Technical Details

The backbone is a compact spatial transformer: 4 layers, hidden size 512, 4 attention heads, GeLU activations, and 0.1 dropout. Spot images are embedded by Prov-GigaPath into 1,536-dimensional features; the model is trained on 928 slides (5% held out for validation) with each whole slide treated as a single training instance, using Adam at a learning rate of 5x10^-4, mean squared error loss, gradient clipping at 1.0, and early stopping. Masking is governed by three sampling schedules — uniform sampling over slide regions, a Beta distribution over masked spots, and an alternation between highly variable genes and the full gene set.

Evaluation spans 6 tasks across 23 datasets and 14 biomarkers. On HEST-Bench, evaluated zero-shot rather than with the benchmark's per-fold retraining, STPath ranks first in Pearson correlation on 8 of 10 datasets and first in adjusted mutual information on 9 of 10, beating the next-best method by 6.9% and 14.4% respectively (two-sided Wilcoxon, P < 0.001). Per-dataset correlations reach 0.75 on IDC and 0.60 on SKCM. Predicted expression of GATA3 and MYBPC1 in breast cancer correlates with ground truth above 0.95, and predicted TP53 expression discriminates mutation status at AUC 0.73 with no supervised training. The authors attribute the model's weakest cases, such as clear cell renal cell carcinoma, to scarcity of relevant data in the pretraining set.

#Applications

STPath lets a pathology lab derive a spatial molecular readout from slides that were never sequenced, which is useful for retrospective cohort analysis, biomarker discovery, and tumor microenvironment characterization at a scale that direct spatial transcriptomics cannot reach. Inferred expression clusters align with pathologist annotations of brain layers, tumor cores, stromal compartments, and immune infiltration zones, and can be appended to visual features as an extra channel for prognostic modeling. Groups running spatial transcriptomics can also use the prompting mode to impute unmeasured spots from a sparse experimental panel.

#Impact

Published in npj Digital Medicine after preprint release in April 2025, STPath is one of the first spatial-transcriptomics predictors to be trained once and applied across organs without per-dataset adaptation, and its zero-shot HEST-Bench protocol is a more realistic evaluation than the leave-one-out retraining the benchmark originally specified. Code and pretrained weights are publicly available, though the repository carries no explicit license file. Performance still tracks pretraining coverage, so organs and platforms that are rare in HEST-1K and STImage-1K4M remain the model's weakest cases.

At a glance

Released
April 2025
Category
Pathology
Organizations
Northeastern University / Yale University / Broad Institute

Links

GitHub RepositoryResearch PaperHuggingFace Model

Tags

foundation_modelgene_expressiongenerativehistologyspatial_transcriptomicssurvival_predictiontransformerzero_shot

Something wrong?

Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.