AnakinHuang/brainscope-scgpt-disease

BrainScope Disease Fine-Tuned scGPT

This repository contains our disease-adapted fine-tuned scGPT model for brain single-cell / single-nucleus RNA-seq analysis.

It is intended to be used with the companion pipeline repository:

• Pipeline repo: YOUR_USERNAME/brainscope-scgpt-pipeline

---

Model summary

This model starts from the original scGPT backbone and is then fine-tuned on disease-related brain single-cell / single-nucleus RNA-seq data for downstream annotation workflows.

This packaged release is intended for:

• disease-aware cell-type annotation
• embedding generation
• comparison against the original scGPT baseline
• downstream error analysis and reproducible model sharing

---

Data context

This release is associated with workflows built on:

• LIBD for smaller pilot experiments and rapid iteration
• BrainScope for larger-scale disease-focused fine-tuning and evaluation

The goal of this model family is to improve robustness on disease-altered cell states relative to healthy-only baselines.

---

Included files

Typical contents of this repository include:

• model.pt
• config.yaml
• preprocessing.json
• vocab.json
• label_map.json
• metrics.json
• requirements.txt
• inference.py
• small example input / output files

---

Intended use

This model is intended for:

• disease-aware annotation of sc/snRNA-seq data
• controlled comparisons with the original scGPT baseline
• reproducible research workflows on brain disease datasets

This release is for research use only and is not a clinical model.

---

Example usage

Download with the Hub

from huggingface_hub import snapshot_download

repo_dir = snapshot_download("YOUR_USERNAME/brainscope-scgpt-disease")
print(repo_dir)

Run through the pipeline

python -m brainscope_scgpt annotate   --input data/query.h5ad   --model-repo YOUR_USERNAME/brainscope-scgpt-disease   --output results/query_annotated.h5ad   --mode small

Large dataset mode:

python -m brainscope_scgpt annotate   --input data/brainscope_full.h5ad   --model-repo YOUR_USERNAME/brainscope-scgpt-disease   --output results/brainscope_full_annotated.h5ad   --mode large

---

Evaluation

Please fill in the exact benchmark numbers you want visible in the public model card.

Suggested structure:

Main metrics

• Accuracy:
• Precision:
• Recall:
• Macro F1:

Benchmark setting

• Train / validation / test split:
• Label space:
• Small or large mode:
• Any freeze / unfreeze strategy:
• Whether MoE was used:

If this release corresponds specifically to one of your final selected models, state that explicitly here.

---

Comparison to the original baseline

This model is intended to be compared against:

• original scGPT baseline
• MoE-enhanced variants
• alternative architectures such as Mamba-based approaches

Suggested points to summarize here after finalization:

• which cell types improved the most
• which confusions remained
• whether disease-aware fine-tuning improved performance on disease-shifted cells

---

Limitations

• Performance depends on preprocessing consistency and gene-vocabulary alignment.
• Performance may change if label definitions differ across datasets.
• This repository does not include all large intermediate artifacts used during training.
• Reference mapping still depends on an external FAISS index if you use the RM workflow.
• This is a research model and not validated for clinical use.

---

Citation

Please cite the original scGPT paper and your project paper when available.

@article{cui2024scgpt,
  title={scGPT: toward building a foundation model for single-cell multi-omics using generative AI},
  author={Cui, Haotian and Wang, Chloe and Maan, Hassaan and others},
  journal={Nature Methods},
  year={2024}
}

---

Contact

Yuesong Huang
University of Rochester
Email: yhu116@ur.rochester.edu