ProteinGPT is a multi-modal protein chat system that allows users to upload protein sequences and/or structures for comprehensive protein analysis and interactive inquiries. It integrates protein sequence and structure encoders with a large language model (LLM) to generate contextually relevant and precise responses.
- Multimodal Input: Accepts protein sequences (FASTA) and structures (PDB).
- Advanced Encoding: Utilizes pre-trained sequence and structure encoders for feature extraction.
- LLM Integration: Aligns protein representations with large language models for insightful responses.
- Extensive Training Data: Trained on a large-scale dataset of 132,092 proteins with annotations and optimized using instruction tuning with GPT-4o.
- Modality Alignment & Instruction Tuning: Two-stage training for enhanced understanding and response accuracy.
Ensure you have the following dependencies installed:
- Python 3.8+
- PyTorch
- CUDA (2 NVIDIA H100 80GB preferred)
- Conda
- Clone the repository:
git clone https://github.com/ProteinGPT/ProteinGPT cd ProteinGPT - Create and activate the environment:
conda env create -f environment.yml conda activate proteingpt
- Install dependencies:
pip install -r requirements.txt
To train ProteinGPT, run the following command:
bash train.shThis script:
- Sets the base model to
Meta-Llama-3-8B-Instruct-hf. - Creates a directory for fine-tuning outputs.
- Launches training using
protein_gpt.pywithconfigs/train_instruction_tuning.yaml.
To test the model interactively, run:
bash src/demo.shOr, to use the protein encoder module:
python demo_esm.py --cfg-path configs/evaluation.yaml --gpu-id 0 --pdb $PATH_TO_STRUCTURE_EMBEDDING --seq $PATH_TO_SEQUENCE_EMBEDDINGProteinGPT/
├── train.sh # Training script
├── environment.yml # Conda environment configuration
├── requirements.txt # Python dependencies
├── src/
│ ├── protein_gpt.py # Main model script
│ ├── demo.sh # Demo script
│ ├── demo_esm.py # Protein encoder demo
│ ├── esm/ # Sequence encoder module
│ ├── prompts/ # Prompt templates
│ ├── data/ # Training datasets
│ ├── configs/ # Configuration files
│ ├── minigpt4/ # Model-related files
├── LICENSE # License file
├── README.md # Guidance
ProteinGPT consists of:
- Protein Sequence Encoder: Uses ESM-2 (
esm2_t36_3B_UR50D). - Protein Structure Encoder: Uses an inverse folding model (
esm_if1_gvp4_t16_142M_UR50). - Projection Layer: Aligns embeddings with the LLM.
- LLM Backbone: Fine-tuned via instruction tuning.
ProteinGPT is trained on the ProteinQA dataset, derived from RCSB-PDB, containing:
- 132,092 protein entries
- Detailed annotations
- 20-30 property tags per protein
- 5-10 curated QA pairs
ProteinGPT outperforms standard LLMs (GPT-4, LLaMA, Mistral) in protein-related tasks, achieving higher semantic similarity scores and improved factual accuracy in closed-ended QA experiments.
- Hallucination Reduction: Improve factual accuracy in responses.
- Citations & References: Integrate verifiable sources.
- Expanded Training Data: Incorporate additional protein knowledge bases.
@article{xiao2024proteingpt,
title={Proteingpt: Multimodal llm for protein property prediction and structure understanding},
author={Xiao, Yijia and Sun, Edward and Jin, Yiqiao and Wang, Qifan and Wang, Wei},
journal={arXiv preprint arXiv:2408.11363},
year={2024}
}