scGPT aims to build a foundation model for single-cell multi-omics analysis using generative AI. It provides pre-trained models and tools for tasks like cell embedding, annotation, and reference mapping, targeting researchers and bioinformaticians working with large-scale single-cell datasets.

How It Works

scGPT leverages a generative transformer architecture, similar to large language models, to learn representations from single-cell data. It processes gene expression profiles as sequences, enabling it to perform various downstream tasks through fine-tuning or zero-shot learning. The model's design allows for efficient handling of large datasets and supports flexible integration with existing bioinformatics tools.

Quick Start & Requirements

• Install via pip: pip install scgpt "flash-attn<1.0.5" (or pip install scgpt "flash-attn<1.0.5" "orbax<0.1.8" if encountering orbax issues).
• Recommended: Python >= 3.7.13, R >= 3.6.1.
• Optional: pip install wandb for logging.
• Flash-attention dependency requires specific GPU and CUDA versions (recommend CUDA 11.7 and flash-attn<1.0.5 as of May 2023).
• Pre-trained checkpoints are available for download, with whole-human recommended.
• Tutorials and online apps are available for reference mapping, cell annotation, and GRN inference.

Highlighted Details

• Pre-trained on over 33 million human cells (whole-human model).
• Supports zero-shot cell embedding and reference mapping to millions of cells efficiently (e.g., 33M cells index < 1GB, search < 1s on GPU).
• Online apps available for browser-based interaction.
• Flash-attention is now an optional dependency, allowing CPU loading.

Maintenance & Community

• Active development with recent updates (Feb 2024) including preliminary HuggingFace integration.
• Tutorials for zero-shot applications and continual pre-trained models are available.
• Contributions are welcomed via pull requests.

Licensing & Compatibility

• License details are not explicitly stated in the README.
• Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

• The README does not explicitly state the license, which is crucial for commercial adoption.
• Flash-attention installation can be complex and requires specific hardware/software configurations.
• Some features, like pretraining code with generative attention masking and HuggingFace integration, are still under development or in preliminary stages.