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 modelTermsPrivacyContact
© 2026 Pulsatance. All rights reserved. ~
Built by Pulsatance
models / pathology / path2space
PathologySpatial omics
National Cancer InstituteCedars-Sinai Medical CenterUniversity of CambridgeUniversity of PennsylvaniaUniversity of SydneySungkyunkwan UniversityUniversity of Maryland, College ParkMedical University of GrazTel Aviv UniversityHebrew University of JerusalemReleased October 2024

Path2Space

Spatial transcriptomics prediction from H&E slides, inferring spot-level expression and tumor microenvironment composition in breast cancer.

The short version

  • —Reads a routine diagnostic slide and returns expression across roughly 14,000 genes
  • —Infers cell-type abundances in the tumor microenvironment without any sequencing
  • —Ensembling 154 checkpoints over frozen pathology features stabilizes per-spot calls
  • —Spatially defined tumor subgroups derived from its output separate by patient survival
  • —Predicts chemotherapy and trastuzumab response from the inferred spatial landscape
50Openness

Where to run it

No providers recorded yet. Browse all providers

Spatial transcriptomics has changed how tumors are understood, replacing a single averaged expression profile with a map of which programs run where — which regions are proliferating, where immune cells have infiltrated, how the stromal boundary behaves. The problem is cost. A spatial assay runs into thousands of dollars per slide, which is why published spatial cohorts number in the tens or low hundreds of patients, far too small for the biomarker discovery that motivates the work. Meanwhile, hematoxylin-and-eosin sections exist for essentially every tumor ever resected, and tens of thousands are already digitized.

Path2Space, from the Cancer Data Science Laboratory at the National Cancer Institute with collaborators across a dozen institutions, closes that gap for breast cancer by predicting spot-level spatial expression directly from the H&E image. Trained on breast cancer spatial transcriptomics cohorts, it emits expression for roughly 14,000 genes at each tile of a whole-slide image and, from that inferred map, estimates the cell-type composition of the tumor microenvironment. First posted as a preprint in October 2024 by Eldad Shulman, Danh-Tai Hoang, Eytan Ruppin and colleagues, it was published in Cell in 2026.

It belongs to the same lineage as Path2Omics, the group's bulk expression and methylation predictor, but answers a different question: not what the tumor expresses on average, but where. What separates it from the crowded field of image-to-expression models is external validation at scale — the trained model was applied unchanged to 976 TCGA breast tumors, a cohort with no spatial data of its own.

#Key Features

  • Spot-level expression across the transcriptome: Roughly 14,000 genes are predicted per tile rather than a curated panel of a few hundred, so the output supports pathway- and signature-level analysis, not just marker lookup.
  • Tumor microenvironment deconvolution: Cell-type abundances are inferred from the predicted spatial expression, giving an immune and stromal composition readout from an image alone.
  • Frozen pathology encoder plus a trained ensemble: A fixed CTransPath tile encoder supplies features and only lightweight regressors are trained, the same adaptation recipe used with slide-level foundation models such as UNI and Prov-GigaPath.
  • Ensemble over 154 checkpoints: Predictions average a nested 22-by-7 fold ensemble rather than relying on a single fit, which damps the fold-to-fold variance that makes single-model spatial predictors unstable.
  • Layered public release: An Apache-2.0 companion repository scores a slide with weights fetched from Zenodo, a sibling repository adds quality-control, deconvolution, treatment-response, and benchmarking code, and a Zenodo codebase archive carries the training scripts.

#Technical Details

Whole-slide images are tiled at 224 pixels and each tile passed through CTransPath, a Swin Transformer pathology encoder pretrained on large histology corpora, producing a 768-dimensional feature vector. Those features feed a nested cross-validation ensemble of two-layer ReLU multilayer perceptrons — 22 outer folds by 7 inner folds, 154 checkpoints in total — whose averaged output is the predicted expression vector for that tile. The released weight package is roughly 6.7 GB, dominated by the MLP ensemble, and the pipeline exposes two entry points: a grid mode that scores a regular tiling of an arbitrary slide, and a spots mode that scores the specific coordinates of a Visium capture array.

Against 21 established spatial expression predictors, Path2Space reported the strongest performance; the preprint quantified this as over 4,300 genes robustly predicted in independent validation. Charting the microenvironment of 976 TCGA breast tumors, the inferred landscapes resolved three spatially defined subgroups with distinct survival outcomes. The clinically pointed result is that microenvironment features derived from these predicted maps outperformed established bulk-sequencing biomarkers at predicting response to chemotherapy and trastuzumab — the cheap image-derived signal beating the expensive assay it approximates.

#Applications

The immediate application is retrospective cohort analysis. Archival breast cancer slide collections with long-term outcome data, most never molecularly profiled, can be given inferred spatial landscapes and mined for treatment-response and prognostic biomarkers at cohort sizes spatial assays cannot reach. For translational groups, the microenvironment readout offers immune-context stratification from material already collected at diagnosis. For spatial biology labs, it serves as a cheap pre-screen, indicating which specimens have heterogeneity worth spending a real assay on.

#Impact

Path2Space's contribution is the demonstration that predicted spatial expression is good enough to carry a clinical argument rather than only to correlate with ground truth — outperforming sequencing-based biomarkers on therapy response is a stronger claim than any correlation coefficient. Its evaluation design also matters, since image-to-expression models are commonly reported on within-cohort cross-validation splits that flatter them, and applying a frozen model to a separate multi-institution cohort is the harder test. The limits are equally clear: training and validation are confined to breast cancer, so cross-indication transfer is unestablished; predictions are statistical inferences, not measurements, and accuracy varies sharply by gene; and the weights and archived codebase carry CC BY-NC 4.0 terms rather than the repository's Apache-2.0, so commercial use needs the authors' authorization.

At a glance

Released
October 2024
Category
Pathology
Organizations
National Cancer Institute / Cedars-Sinai Medical Center / University of Cambridge / University of Pennsylvania / University of Sydney / Sungkyunkwan University / University of Maryland, College Park / Medical University of Graz / Tel Aviv University / Hebrew University of Jerusalem

Links

GitHub RepositoryGitHub RepositoryResearch PaperbioRxiv PreprintDatasetLink

Tags

biomarker_discoverycancergene_expression_predictionhistologymultilayer_perceptronspatial_transcriptomicstransfer_learningvision_transformer

Something wrong?

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