OmicsML/dance

What it solves

DANCE is a deep learning toolkit designed to standardize and optimize the analysis of single-cell gene expression data. It addresses two primary challenges: the difficulty of reproducing benchmarks across different computational methods (DANCE 1.0) and the trial-and-error nature of data preprocessing, where the optimal combination of normalization and dimensionality reduction often varies by dataset and model (DANCE 2.0).

How it works

The project is divided into two main versions:

• DANCE 1.0 acts as a benchmark platform. It provides a unified Python package that implements popular single-cell analysis methods and includes tools for data downloading, processing, and evaluation to ensure experiments can be reproduced faithfully.
• DANCE 2.0 introduces an automated preprocessing recommendation platform. It uses a Method-Aware Preprocessing (MAP) module to find the best pipeline for a specific method and a Dataset-Aware Preprocessing (DAP) web service to provide instant recommendations for new datasets.

Who it’s for

Computational biologists and researchers working with single-cell omics data who need to implement deep learning models for cell type annotation, clustering, or gene imputation.

Highlights

• Comprehensive Modules: Supports single-modality analysis (annotation, clustering, imputation), multimodal omics (joint embedding, modality prediction/matching), and spatially resolved transcriptomics.
• Automated Preprocessing: Moves beyond manual trial-and-error with data-driven pipeline recommendations.
• Unified Framework: Simplifies the installation and execution of diverse algorithms (e.g., GNNs, Autoencoders, GANs) through a single library.
• Benchmarking Tools: Provides ready-to-use example scripts to reproduce original paper results.