scTrilemma

Released checkpoint of scTrilemma: Balancing Identity, Invariance, and Fidelity in Single-Cell Representation Learning (NeurIPS 2026).

Paper: scTrilemma: Balancing Identity, Invariance, and Fidelity in Single-Cell Representation Learning

• final.ckpt — PyTorch Lightning checkpoint (weights + model/data configuration, 223 MB), SHA-256 58981767d498c5d9c4087276bc2853cb6d5c338fc5048d84b3d8f8c9cacf5f76.
• Trained for 25,000 steps on the 2025-01-30 CELLxGENE Census release (62.6 M cells, 478 datasets).
• Code, training recipe and evaluation: https://github.com/yunhak0/scTrilemma

Usage

git clone https://github.com/yunhak0/scTrilemma && cd scTrilemma
pixi install --locked
bash scripts/download_checkpoint.sh

from huggingface_hub import hf_hub_download
path = hf_hub_download("yunhak0/scTrilemma", "final.ckpt")

See the GitHub README for the inference API (embed_adata, reconstruct_adata).

Citation

@inproceedings{oh2026sctrilemma,
  title={scTrilemma: Balancing Identity, Invariance, and Fidelity in Single-Cell Representation Learning},
  author={Oh, Yunhak and Lee, Yoonho and Lee, Junseok and Lee, Namkyeong and Hwang, Sang-Yeon and Piao, Yinhua and Kim, Hyomin and Kim, Seonghwan and Lim, Jaechang and Kim, Woo Youn and Ahn, Sungsoo and Park, Chanyoung},
  booktitle={The Fortieth Annual Conference on Neural Information Processing Systems},
  year={2026},
  url={https://arxiv.org/abs/2609.38840}
}