scGenAI

Author: Ruijia Wang

Date: 2024-10-14

Version: 1.0.0

About scGenAI

---

scGenAI is a Python package for single-cell RNA sequencing (scRNA-seq) data prediction and analysis using large language models (LLMs). The package allows users to train, fine-tune, and make predictions on single-cell data using transformer-based models, including custom versions of LLaMA, GPT, BigBird, and scGenT. It provides multi-GPU support with PyTorch DistributedDataParallel (DDP).

Table of Contents

---

• I. Installation
• II. Quick Start/Tutorials
  • Option 1. Use scGenAI through notebook or python script
  • Option 2. Use scGenAI through Command Line Interface
• III. Documentation
• VI. License

I. Installation

---

To install the package, use the following steps:

0. Optional: Create a Env for scGenAI:

   conda create -n scGenAI python==3.10
   conda activate scGenAI

1. Clone the repository:

   git clone https://github.com/VOR-Quantitative-Biology/scGenAI.git

2. Navigate to the project directory:

   cd scGenAI

3. Install dependencies, then install scGenAI:

   pip install -r requirements.txt
   pip install .

II. Quick Start/Tutorials

---

Once installed, scGenAI can be accessed through either (1) python IDE or notebook or (2) the command line interface (CLI). You can train, predict, or fine-tune a model by calling the scGenAI in python or CLI commands along with a configuration YAML file containing your settings.

Option 1. Use scGenAI through notebook or python script

As a quick start, we highly recommend users to begin with the following tutorials using a testing size of data (40 cells) and config template files according to the training/prediction purposes:

Testing Case	Tutorial	Training/Finetune Config Template	Prediction Config Template
Modeling healthy cell type	TrainData_Tutorial	config_Train	config_Prediction
Modeling disease/cancer cell type	TrainData_Tutorial	config_Train or config_Train	config_Prediction
Modeling cell genotype	TrainData_Tutorial	config_Train	config_Prediction
Modeling using multiomics data	TrainData_Tutorial	config_Train	config_Prediction
Fine-tune using pretrained model	FinetuneData_Tutorial	config_Train	config_Prediction

In addition to the testing data, we also provide full-size datasets and config files according to the training/prediction purposes.

Study	Tutorial	Training/Finetune Config	Prediction Config
Mouse eye cell types gene list context	TrainData_Tutorial	config_Train	config_Prediction
Myeloma myeloid cell types genomic context	TrainData_Tutorial	config_Train	config_Prediction
AML cell types and cell status biofunction context	TrainData_Tutorial	config_Train	config_Prediction
PBMC CITE-Seq	TrainData_Tutorial	config_Train	config_prediction
Fine-tune of bone marrow cell types	FinetuneData_Tutorial	config_Finetune	config_Prediction

Option 2. Use scGenAI through Command Line Interface

The CLI supports the following commands:

• Train a model:

  scgenai train --config_file <path_to_config.yaml>

• Make predictions:

  scgenai predict --config_file <path_to_config.yaml>

• Fine-tune a pre-trained model:

  scgenai finetune --config_file <path_to_config.yaml>

III. Documentation

---

Please see the full documentation for the details usage of scGenAI.

• Input Files and parameters
• Configuration
• Run scGenAI
• Output Files

V. License Agreement

---

• The use of scGenAI is governed by a custom license permitting non-commercial use only LICENSE. This package is freely available to individuals, universities, non-profit organizations, educational institutions, and government entities for non-commercial research or journalistic purposes.

• By cloning or downloading this repository, the user acknowledge that the user has read, understood, and agree to abide by the terms outlined in the LICENSE file.