PANDA — Pan-tissue Adversarial Normalized Domain-invariant Anchored MLP

Prototype-anchored MLP classifier for scRNA-seq cell identity across skin, hematopoietic, and pancreatic tissues. Trained under a composite of SupCon + VICReg + prototype-InfoNCE + GRL dataset+depth adversary + HSIC decorrelation + prototype-repulsion.

Two variants: PANDA-PCA and PANDA-Marker (adds a marker gene channel).

Code + paper: https://github.com/bryanc5864/PRISM

Contents

Path	Description
checkpoints/{system}/{pca,marker}/panda_final.pt	Final trained weights per system × variant (6 core models)
checkpoints/pan_skin_dingwall_derm/panda_final.pt	Line C: PANDA-Marker trained on Dingwall Derm labels
data/corpus/{system}/harmonized/	Training corpora (h5ad + stats + PCA basis)
data/external_labels/	Paper-supplement label files per source study
data/processed/dingwall_replica/	Independent scanpy reproduction of Dingwall Seurat pipeline
discovery/	Discovery-analysis outputs backing every paper claim
figures/	Main + supplement + biology figures + merged PDFs
panda/, scripts/	Model + analysis code (also on GitHub)
PAPER.tex, PAPER.pdf	Manuscript
README.md	Full end-to-end reproduction recipe

Quick fetch

# essentials only (~30 GB)
huggingface-cli download bryan7264/PANDA \
    --local-dir . \
    --include "checkpoints/**" "data/corpus/**" "data/external_labels/**" "discovery/**"

# individual system
huggingface-cli download bryan7264/PANDA \
    --local-dir . \
    --include "data/corpus/pan_skin/**" "checkpoints/pan_skin/**"

Usage

import torch
from panda.model import PANDAEncoder

ck = torch.load("checkpoints/pan_skin/marker/panda_final.pt", map_location="cpu",
                weights_only=False)
model = PANDAEncoder(variant="marker", n_pca=50,
                     n_markers=len(ck["marker_genes"]),
                     n_classes=len(ck["classes"]), n_sub=3,
                     n_datasets=len(ck["datasets"]))
model.load_state_dict(ck["model"])
model.eval()

See PAPER.pdf for full experimental setup and README.md for the reproduction recipe.