This is the official repository for the paper EvoLlama: Enhancing LLMs' Understanding of Proteins via Multimodal Structure and Sequence Representations.

[Dataset] | [Model] | [Preprint]

We recommend using Python >= 3.9, and then simply use pip to install the required packages:

pip install -r requirements.txt

Model weights are publicly available on 🤗HuggingFace. During training, the parameters of Llama-3 are frozen. To initialize EvoLlama, you need to manually download the LLM weights at meta-llama/Meta-Llama-3-8B-Instruct.

For projection-tuned EvoLlama, only the projection layers are trainable. Therefore, you need to manually download the ProteinMPNN weights/ GearNet weights, and the ESM-2 weights.

The table below provides a summary of the EvoLlama model family and includes links to their model weights on 🤗HuggingFace.

Helper functions for initializing EvoLlama and generating responses are provided in src/infer/infer.py. Note that the function infer() accepts a list of lists of PDB files and sequences, and a list of arbitrary prompts as inputs. When utilizing EvoLlama without a structure/sequence encoder, set the corresponding parameter None.

import os
from src.infer.infer import init_evo_llama, infer
# 1. Initialize EvoLlama
model_weights_path = '/path/to/EvoLlama'
llm_weights_path = '/path/to/llm'
evo_llama = init_evo_llama(
    structure_encoder_path=os.path.join(model_weights_path, 'structure_encoder_weights'),
    structure_encoder_name='ProteinMPNN',
    sequence_encoder_path=os.path.join(model_weights_path, 'sequence_encoder'),
    llm_path=llm_weights_path,
    projection_path=os.path.join(model_weights_path, 'projection_weights.bin'),
    projection_fusion=True,
    is_inference=True
)
# 2. Inference with EvoLlama
pdb_files = ['examples/ea91f233142ab1a17749be765a461255.pdb']  # We use the MD5 hash of the protein sequence as the filename.
sequences = ['MANHKSTQKSIRQDQKRNLINKSRKSNVKTFLKRVTLAINAGDKKVASEALSAAHSKLAKAANKGIYKLNTVSRKVSRLSRKIKQLEDKI']
prompt = 'Analyze the given amino acid sequence, and determine the function of the resulting protein, its subcellular localization, and any biological processes it may be part of.'
responses = infer(evo_llama, [pdb_files], [sequences], [prompt])

Additionally, simply run scripts scripts/eval_molinst.sh and scripts/eval_peer.sh to evaluate EvoLlama on the protein understanding and protein property prediction tasks, respectively.

# Evaluate EvoLlama on the protein understanding tasks.
bash scripts/eval_molinst.sh
# Evaluate EvoLlama on the protein property prediction tasks.
bash scripts/eval_peer.sh

Coming soon ...

@misc{liu2024evollama,
    title={EvoLlama: Enhancing LLMs' Understanding of Proteins via Multimodal Structure and Sequence Representations}, 
    author={Nuowei Liu and Changzhi Sun and Tao Ji and Junfeng Tian and Jianxin Tang and Yuanbin Wu and Man Lan},
    year={2024},
    eprint={2412.11618},
    archivePrefix={arXiv},
    primaryClass={cs.LG},
    url={https://arxiv.org/abs/2412.11618}, 
}