mims-harvard/ProCyon-Instruct

This repository contains the ProCyon-Instruct used to train the ProCyon family of models. Please see installation instructions on our GitHub repo for details on how to configure the dataset for use with pre-trained ProCyon models. For additional technical details, please refer to our overview page or the paper.

The repository contains three top-level directories:

• integrated_data/v1 - The primary component of the dataset: the amino acid sequences and associated phenotypes used for constructing instruction tuning examples.
• generated_data - Contains additonal artifacts beyond amino acids and phenotypes. Generated by the ProCyon team and used for model training and evaluation.
• model_weights - Contains pre-trained model weights used for initializing ProCyon models. Note that the model weights themselves are not contained in this repository but rather are expected to be downloaded here from their respective repositories.

Within integrated_data, there are four main types of directories:

• {amino_acid_seq_type} - directories containing information for amino acid sequences themselves, where amino_acid_seq_type is one of ["domain", "peptide", "protein"]. Each directory contains the following files:
  • {amino_acid_seq_type}_sequences.fa - FASTA file containing the raw amino acid sequence for each entity
  • {amino_acid_seq_type}_info_filtered.pkl - Pickled Pandas DataFrame containing the mapping from the amino acid sequence's database ID (e.g. UniProt ID for proteins) to a numeric index used within ProCyon-Instruct. Two columns:
    • index - numeric ID within ProCyon-Instruct
    • {amino_acid_seq_type}_id - ID within original database
• {phenotype_type} - directories containing information for each phenotype entity. Each directory contains the following files:
  • {phenotype_type}_info_filtered.pkl - Pickled Pandas DataFrame containing mapping from phenotype's database ID to numeric ID within ProCyon-Instruct, and various textual descriptions within each database. Has the following columns:
    • index - numeric ID within ProCyon-Instruct
    • {phenotype_type}_id - ID within original database
    • additional columns coming from the original databases giving various textual descriptions of the phenotype. Used to create the instruction tuning examples
  • {phenotype_type}_info_filtered_composed.pkl - Pickled Pandas DataFrame containing the same data as {phenotype_type}_info_filtered.pkl but with additional columns giving compositions of individual text columns from the original DataFrame.
• {amino_acid_seq_type}_{phenotype_type} - directories containing information on the associations between amino acid sequences and phenotypes. Each directory contains a subdirectory named based on the method used for generating dataset splits within that database. Please see the methods section of our manuscript for more details. Within these subdirectories there are two files:
  • {amino_acid_seq_type}_{phenotype_type}_relations.unified.csv - CSV file containing relations expressed in original database IDs. Contains six columns:
    • text_id - ID from original phenotype database
    • seq_id - ID from original sequence database
    • text_type - largely redundant with phenotype_type, may be helpful when concatenating many assocations files
    • seq_type - largely redundant with amino_acid_seq_type, may be helpful when concatenating many assocations files
    • relation - largely redundant with f{amino_acid_seq_type}_{phenotype_type}, may be helpful when concatenating many assocations files. For some datasets such as DrugBank and GO, this column takes on different values within the same file and expresses distinct relations, e.g. GO molecular function vs GO biological process.
    • split - Assigned data split for this association. CL_train are training associations, CL_val_* are validation associations, and eval_* are test associations. Both CL_val and eval have sufficies indicating whether these relations are zero-shot with respect to the phenotype, where _zero_shot indicates a zero-shot relation, _[num]_shot indicates a few-shot relation, and _pt_ft indicates relations where the phenotype is seen frequently in training.
  • {amino_acid_seq_type}_{phenotype_type}_relations_indexed.unified.csv - Identical to the above CSV file, but with relations expressed using ProCyon internal numeric IDs.
• {amino_acid_seq_type}_{amino_acid_seq_type} - directories containing information on the associations between two amino acid sequences, e.g. protein-protein interactions. Format is largely the same as above except with seq_id_1 and seq_id_2 columns instead of seq_id and text_id