geneformer-oa-myeloid

Geneformer-OA-Myeloid

Geneformer-OA-Myeloid is a released collection of two Geneformer-based downstream classifiers for osteoarthritis synovial myeloid single-cell analysis.

Released Models

Performance

Disease classifier

• Base model: Geneformer-V2-104M
• Internal held-out accuracy: 0.9961
• Internal held-out macro-F1: 0.9956
• Internal held-out ROC-AUC: 0.99997

State classifier

• Base model: Geneformer-V2-104M
• Internal held-out accuracy: 0.9507
• Internal held-out macro-F1: 0.8985

External validation

• Cohort: GSE253198
• OA/Normal transfer accuracy: 0.8146
• OA/Normal transfer macro-F1: 0.4704
• OA ROC-AUC: 0.5184

The disease classifier performs strongly inside the source synovial cohort but shows limited cross-cohort transfer, so this release should be used as a research model rather than a cohort-invariant clinical classifier.

Repository Contents

models/
  disease_classifier/
  state_classifier/
results/
  external_validation/
scripts/

Usage

This release distributes model metadata, evaluation summaries, and helper scripts through GitHub, while large binary weights are intended for Hugging Face or Git LFS distribution.

Minimal loading pattern:

from pathlib import Path
from transformers import BertForSequenceClassification

model_dir = Path("models/disease_classifier")
model = BertForSequenceClassification.from_pretrained(model_dir)

Scope

• Species: human
• Tissue focus: synovium-derived OA myeloid cells
• Intended use: representation learning, disease-state scoring, and method development in related OA synovial datasets
• Not intended for: diagnosis, prognosis, treatment selection, or unsupported cross-tissue deployment

Data Provenance

This release does not redistribute raw single-cell count matrices. The released models and summaries were generated from public datasets including:

• GSE216651
• GSE152805
• GSE253198

License and Attribution

This release is distributed under Apache License 2.0 as a derivative downstream work of ctheodoris/Geneformer.

• Upstream foundation model: ctheodoris/Geneformer
• Proper description: A Geneformer-based OA myeloid fine-tuned model collection