ProteinLMDataset & ProteinLMBenchmark

Paper Link: https://arxiv.org/abs/2406.05540

Paper Presentation: Youtube: https://www.youtube.com/watch?v=2EXWnSHyU0U

ProteinLMDataset SSL Data Generation

Biology Chinese/English text pair subset

Our data source comprises several Chinese journal websites (e.g., Chinese Journal of Biotechnology, Chinese Journal of Cell Biology), which provide both Chinese and English abstracts of their published articles. We selected journals and articles in the field of protein engineering and extracted these Chinese and English abstracts to form the dataset. You may attempt to scrape data from other websites as well, but you must first pay attention to their usage agreements.

PubMed abstracts subset

You can obtain abstracts for PubMed articles through various methods available on their official website: https://pubmed.ncbi.nlm.nih.gov/download/

PMC full-text manuscripts subset

1. Preperation: You can download the original PMC dataset from relevant NCBI websites: https://www.ncbi.nlm.nih.gov/pmc/tools/ftp/
2. For the next, you can use code in ssl/pmc to get the articles that may related to protein knowledges.
3. For a list of articles highly relevant to protein topics, you can download the corresponding file pdb_pubmed.csv from the Structure Integration with Function, Taxonomy and Sequence (SIFTS) official FTP server( https://ftp.ebi.ac.uk/pub/databases/msd/sifts/ ). After placing it in the appropriate location, you can execute ssl/pmc/highly_pmc_1_pmid_pmcid.py to obtain the file pubmed_pmc.csv, which contains all the available PMC IDs.
4. For a list of articles not directly related to protein topics, you can download the corresponding file pdb_chain_uniprot.csv from the official SIFTS FTP server. After placing it in the appropriate location, execute all the files in the ssl/pmc folder sequentially according to the numbers following week_pmc_. This process will extract the articles and save them in the Data/Literature/articles_json folder.
   Note: You were advised to place the downloaded original PMC data in the /Data/paper_text_dataset/pmc folder to save time.

Protein sequence/PubMed abstracts subset

You can generate an embedded text description of the protein sequence dataset step-by-step using the code in ssl/seq_in_text. The complete workflow diagram is as follows:

UniProt-SwissProt subset

Directly downloading our pre-downloaded and pre-processed original datasets proteins1.json and proteins2.json and placing them in the appropriate location, you can directly run the Python files within ssl/Uniprot to obtain the dataset.

ProteinLMDataset SFT Data Generation

UniProt-SwissProt multi project subset

Directly downloading our pre-downloaded and pre-processed original datasets proteins1.json and proteins2.json and placing them in the appropriate location, you can directly run the Python files within sft/uniprot_1_insert_in_template.py to obtain this subset.

Enzyme Chain of Thought(ECoT) subset

1. Preperation: Please visit the Enzyme Nomenclature website of the Nomenclature Committee of the International Union of Biochemistry and Molecular Biology (NC-IUBMB) at https://iubmb.qmul.ac.uk/enzyme/index.html. Locate the "List of Recommended Names for Enzymes" section, download all links in the "separate" column of the table, and store them in the Data/EC_file folder.
2. Run the files beginning with ECoT in numerical order. Finally, you will obtain the Enzyme_CoT.json file, which should be saved in the Data/SFT folder.

ProteinLMBenchmark Part

Benchmark Generation:

Preperation: Before starting, refer to the methods mentioned in the vLLM official documentation (https://github.com/vllm-project/vllm) to wrap the local large model in the OpenAI format interface, making it easier to integrate into the LangChain process. At each stage, we will use the model 'Mixtral-8x7b-Instruct-v0.1'.

Please execute all files beginning with benchmark_{%d} in numerical order. During execution, pay attention to the required benchmark scale. After generation, our benchmark underwent further manual review by experts in the protein field. You can directly download our benchmark dataset from the link: https://huggingface.co/datasets/tsynbio/ProteinLMBench.