Welcome to scDynOmics’s documentation!

scDynOmics is an optimized, pretrainable transformer model designed for representation learning from single-cell multiomics data.

Motivated by gene regulatory networks, scDynOmics adopts a Linformer-style attention mechanism to efficiently scale to coding-genome wide multimodal inputs. By pretraining on paired single-cell transcriptomic and chromatin accessibility profiles, the model yields compact embeddings that represent cellular states and developmental dynamics. scDynOmics also supports unimodal data for pretraining and/or finetuning.

For versatile downstream applications, the framework employs low-rank adaptation modules for parameter-efficient fine-tuning, enabling rapid adaptation to diverse biological tasks.

Note

This project is under active development.

Key Features

  • Scalable Architecture: Utilizes a Linformer-style attention mechanism (via scdynomics.Masked_Pretrainer) to handle large-scale inputs efficiently.

  • Parameter-Efficient Fine-Tuning: Incorporates low-rank adaptation modules (scdynomics.FT_MLP_Classifier, etc) for lightweight adaptation to diverse biological tasks.

  • Interpretable Framework: Deciphers cellular heterogeneity and provides interpretable factors for developmental dynamics.

Citation

If you use scDynOmics in your research, please cite our bioRxiv preprint:

@misc{yu2026scdynomics,
  title={scDynOmics: An Optimized Transformer Model for Representation Learning from Single-Cell Multiomics},
  author={Yu, Gang and Ramnarine, Timothy J. S. and Klughammer, Johanna and Mages, Simon W.},
  year={2026},
  publisher={bioRxiv},
  url={https://doi.org/10.64898/2026.02.28.708160},
  doi={10.64898/2026.02.28.708160}
}

Contents

Getting Started