scmiracle 0.0.2

A torch-based integration method for single-cell multi-omic data.

MIRACLE: A Continual Integration Method for Single-cell Data

By employing dynamic architecture adaptation and data rehearsal strategies , MIRACLE enables continual integration of diverse datasets while preserving biological fidelity over time.

---

MIRACLE , an online integration framework for multimodal single-cell integration via continual learning (CL). CL enables models to incrementally incorporate new data while preserving previously acquired knowledge. MIRACLE employs CL strategies including dynamic architecture adaptation and data rehearsal to enhance adaptability and knowledge retention.

• MIRACLE Documentation: scmiracle.readthedocs.io

✨ Key Features

• Boosted Efficiency with Continual Integration: Incrementally add new data batches to an existing model, which eliminates the need for complete retraining and significantly reduces computational requirements.
• Dynamic Feature Space Adaptation: Seamlessly incorporate new data containing novel features (e.g., new genes or proteins not present in the original dataset). The model dynamically expands its feature space, ensuring no information is lost when integrating data from different technologies or antibody panels.
• Intelligent Data Summarization for Rehearsal: Employs a distribution-preserving sampling strategy (BTS) to compress past data into a small, representative core set. This "rehearsal" dataset enables efficient knowledge retention during updates without needing to store all historical data, making the process highly scalable and shareable.

🚀 Installation

Get started with MIRACLE by setting up a conda environment.

conda create -n scmiracle python=3.12
conda activate scmiracle
pip install scmiracle

⚡ Getting Started

To get started, please refer to our documentation.

📈 Reproducibility

To reproduce the results from our publication, please visit the reproducibility branch of this repository: https://github.com/sc-miracle/miracle-reproducibility/

📝 License

MIRACLE is available under the MIT License.