Request access to Paladin
Access requests are reviewed manually by the Paladin team.
Please provide the information below so the Paladin team can review your request. Your Hugging Face username and email address are shared automatically.
Log in or Sign Up to review the conditions and access this model content.
Paladin
Paladin: A foundation model for pathomics predicts patch-level and slide-level omics directly from a whole-slide H&E image.
- Private canonical code, documentation, and notebooks for authorized users: github.com/zhihuanglab/Paladin
- This gated Hugging Face repository is the canonical home for the Paladin checkpoint and large example assets. File downloads require a manually approved access request. It does not duplicate the source tree.
- Released Paladin-inferred TCGA and CPTAC spatial omics maps: zhihuanglab/Paladin_TCGA_CPTAC_omics
Model assets
weights/Paladin/Paladin.pth
weights/Paladin/Paladin.json
weights/Paladin/omics_names.json
Request access on this model page and wait for approval. After approval, download the assets into a GitHub checkout with:
hf auth login
hf download zhihuanglab/Paladin \
weights/Paladin/Paladin.pth \
weights/Paladin/Paladin.json \
weights/Paladin/omics_names.json \
--local-dir .
Paladin uses the gated Bioptimus H-optimus-0 encoder; that encoder is not redistributed here.
TCGA and CPTAC spatial omics maps
Ready-to-use .Paladin.h5 predictions are distributed through the
Paladin_TCGA_CPTAC_omics dataset.
The released files can be analyzed directly without rerunning WSI inference.
hf download zhihuanglab/Paladin_TCGA_CPTAC_omics \
--repo-type dataset \
--local-dir Paladin_TCGA_CPTAC_omics
Select a cohort with --include, for example
--include "CPTAC/cptac_coad/*.Paladin.h5", to avoid downloading the full
collection.
TCGA-COAD example
The example uses the same representative COAD slide selected by the reference CNV analysis:
examples/tcga_coad/
├── TCGA-5M-AATE-01Z-00-DX1.483FFD2F-61A1-477E-8F94-157383803FC7.svs
├── TCGA-5M-AATE-01Z-00-DX1.483FFD2F-61A1-477E-8F94-157383803FC7.Paladin.h5
├── TCGA-5M-AATE-01Z-00-DX1.483FFD2F-61A1-477E-8F94-157383803FC7.tumor_annotation.npz
├── tumor_labels_224px.json
└── manifest.json
The executed GitHub notebook
examples/TCGA_COAD_CNV_demo.ipynb
demonstrates inference and an eight-panel visualization of the H&E WSI,
reference tumor mask, CNV burden, TP53 CNV, DNA-methylation burden, CDX2
spatial RNA, hsa-miR-21-5p miRNA, and PanCK CODEX protein.
The tumor mask is an external reference annotation, not a Paladin prediction
and not part of the default output HDF5 schema. The 224 px source labels are
mapped onto Paladin's 507 px patches by majority vote over source-tile centers;
the original JSON and the aligned NPZ are both kept for provenance and
reproducibility. CNV, methylation, and miRNA are bulk-trained heads, so their
patch maps are model-derived spatial predictions and slide_values remains
their primary output. Spatial RNA and CODEX protein are patch-level spatial
heads. All maps show model-space values rather than assay-native units.
The prediction contains 26,147 tissue patches at a 128 µm patch edge. Its portable HDF5 schema is:
/coords [N, 2] int64
/tasks/<task>/patch_values [N, D] float16
/tasks/<task>/feature_names [D] string
/tasks/<task>/slide_values [D] float32
Prediction heads
| Task | Features | Primary level |
|---|---|---|
vaf_bulk |
3,525 | slide |
rnaseq_bulk |
4,048 | slide |
cnv_bulk |
4,033 | slide |
methy_bulk |
7,387 | slide |
mirna_bulk |
743 | slide |
rnaseq_st |
4,048 | patch |
codex_protein |
73 | patch |
Checksums and provenance
Machine-readable sizes, SHA-256 checksums, inference parameters, and QC status
are recorded in examples/tcga_coad/manifest.json. The example WSI originates
from TCGA/GDC. Its upstream terms remain applicable; the model repository
license does not relicense third-party TCGA data.
Intended use and limitations
Paladin is intended for computational pathology and spatial/multi-omics research. Patch-level outputs from bulk-trained heads are model predictions, not measured spatial assays. Values must not be interpreted automatically as raw counts, integer copy number, calibrated log2 ratios, methylation beta values, or clinical test results. This repository is not a diagnostic medical device and its outputs require independent validation.
License
Paladin software and model weights are provided under the PENN Academic Software License Agreement for Paladin. The complete license text is controlling. Third-party models and source datasets remain subject to their respective upstream terms.
Acknowledgments
We thank the TCGA Research Network and participating patients and institutions, Bioptimus for H-optimus-0, and the Mahmood Lab for HEST.