| ## Introduction | |
| <p align="center"> | |
| <br> | |
| <img src="assets/FAPM.png"/> | |
| <br> | |
| <p> | |
| Huggingface repo: *https://huggingface.co/wenkai/FAPM/* | |
| ## Installation | |
| 1. (Optional) Creating conda environment | |
| ```bash | |
| conda create -n lavis python=3.8 | |
| conda activate lavis | |
| ``` | |
| 2. for development, you may build from source | |
| ```bash | |
| git clone https://github.com/xiangwenkai/FAPM.git | |
| cd FAPM | |
| pip install -e . | |
| # if needed | |
| # pip install Biopython | |
| # pip install fair-esm | |
| ``` | |
| ### Datasets | |
| #### 1.raw dataset | |
| Raw data are avaliable at *https://ftp.uniprot.org/pub/databases/uniprot/previous_releases/release-2023_04/knowledgebase/*, this file is very large and need to be processed to get its name, sequence, GO label, function description and prompt. | |
| The domain level protein dataset we used are avaliable at *https://ftp.ebi.ac.uk/pub/databases/interpro/releases/95.0/protein2ipr.dat.gz* | |
| In this respository, We provide the experimental train/val/test sets of Swiss-Prot, which are avaliable at data/swissprot_exp | |
| #### 2.ESM2 embeddings | |
| Source code for ESM2 embeddings generation: *https://github.com/facebookresearch/esm* | |
| The generation command: | |
| ```bash | |
| conda activate FAPM | |
| python esm_scripts/extract.py esm2_t36_3B_UR50D you_path/protein.fasta you_path_to_save_embedding_files --repr_layers 36 --truncation_seq_length 1024 --include per_tok | |
| ``` | |
| Example: | |
| ``` | |
| conda activate FAPM | |
| python esm_scripts/extract.py esm2_t36_3B_UR50D data/fasta/example.fasta data/emb_esm2_3b --repr_layers 36 --truncation_seq_length 1024 --include per_tok | |
| ``` | |
| The default path to save embedding files is **data/emb_esm2_3b** | |
| You can refer to *data/fasta/prepare_custom_fasta.py* to prepare your custom fasta data. | |
| ## Pretraining language models | |
| Source: *https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B* | |
| ## Training | |
| data config: lavis/configs/datasets/protein/GO_defaults_cap.yaml | |
| stage1 config: lavis/projects/blip2/train/protein_pretrain_stage1.yaml | |
| stage1 training command: run_scripts/blip2/train/protein_pretrain_domain_stage1.sh | |
| stage2 config: lavis/projects/blip2/train/protein_pretrain_stage2.yaml | |
| stage2 training/finetuning command: run_scripts/blip2/train/protein_pretrain_domain_stage2.sh | |
| ## Trained models | |
| The models are avaliable at **https://huggingface.co/wenkai/FAPM/tree/main/model** | |
| You can also download our trained models from google drive: *https://drive.google.com/drive/folders/1aA0eSYxNw3DvrU5GU1Cu-4q2kIxxAGSE?usp=drive_link* | |
| ## Testing | |
| config: lavis/projects/blip2/eval/caption_protein_eval.yaml | |
| command: run_scripts/blip2/eval/eval_cap_protein.sh | |
| ## Inference example | |
| ``` | |
| python FAPM_inference.py \ | |
| --model_path model/checkpoint_mf2.pth \ | |
| --example_path data/emb_esm2_3b/P18281.pt \ | |
| --device cuda \ | |
| --prompt Acanthamoeba \ | |
| --prop True | |
| ``` | |