IRIS pretrained checkpoints
Pretrained SCANVI models for IRIS, which infers signaling-pathway activity from single-cell transcriptomes. These are the checkpoints used by the analyses in the IRIS manuscript (Nature Methods).
Code: https://github.com/Pulin-Li-Lab/IRIS-signaling-inference
Contents
Ten checkpoints, five pathways × two training contexts.
| Checkpoint | Pathway | Trained on | Arch |
|---|---|---|---|
iris_gastrulation_RA_scanvi_h256_z70 |
RA | mouse gastrulation atlas (E6.5–E8.5) + screens | h256 z70 |
iris_gastrulation_Wnt_scanvi_h64_z30 |
WNT | " | h64 z30 |
iris_gastrulation_Fgf_scanvi_h128_z70 |
FGF | " | h128 z70 |
iris_gastrulation_Bmp_scanvi_h256_z70 |
BMP | " | h256 z70 |
iris_gastrulation_TgfB_scanvi_h1024_z30 |
TGFβ | " | h1024 z30 |
iris_foregut-mesenchyme_RA_scanvi_h256_z70 |
RA | E9–E9.5 foregut mesenchyme (Han et al. 2020) + screens | h256 z70 |
iris_foregut-mesenchyme_Wnt_scanvi_h64_z30 |
WNT | " | h64 z30 |
iris_foregut-mesenchyme_Fgf_scanvi_h128_z70 |
FGF | " | h128 z70 |
iris_foregut-mesenchyme_Bmp_scanvi_h256_z70 |
BMP | " | h256 z70 |
iris_foregut-mesenchyme_TgfB_scanvi_h1024_z30 |
TGFβ | " | h1024 z30 |
Each directory is an scvi-tools save directory containing model.pt.
Architectures are per-pathway, as published — not a single shared configuration.
gastrulation checkpoints support the in vivo lineage-dynamics analysis;
foregut-mesenchyme checkpoints support the organ-specific mesenchyme analysis,
whose published result is WNT/BMP enrichment in respiratory mesenchyme.
Environment
Trained and saved with scvi-tools 0.20.3 (torch 2.2.2+cu121, scanpy 1.10.0). scvi-tools 1.x changes the model-loading API and will not load these directly.
Usage
from huggingface_hub import snapshot_download
path = snapshot_download(
repo_id="nhutchins627/iris-pretrained-checkpoints",
allow_patterns="iris_gastrulation_Bmp_scanvi_h256_z70/*",
)
Then load via IRIS:
iris_obj.load_pretrained_model([path], ["Bmp"])
Citation
Cite the IRIS paper. Datasets the models were trained on carry their own
accessions — see reproducibility/docs/DATA.md in the code repository.