PARSE

Protein Annotation by Residue-Specific Enrichment: a site-based statistical method for interpretable protein function annotation

Install dependencies

We recommend installing all dependencies in a conda environment:

conda create -n parse python=3.10

To install, run the following script on a GPU-enabled machine. To install for CPU only, replace cu117 with cpu.

./install_dependencies.sh

Download data

Download CSA reference data and precomputed embeddings for AlphaFoldDB datasets (human and dark proteomes) from Zenodo.

CSA data (required):

• csa_function_sets_nn.pkl: function sets dictionary
• csa_site_db_nn.pkl: embeddings database
• function_score_dists.pkl: function-specific background distributions

AlphaFoldDB precomputed embeddings (optional):

• af2_human_lmdb.zip: AF2 human proteome embeddings in LMDB format
• af2_dark_hernandez_lmdb.zip: AF2 dark proteome embeddings from Barrio-Hernandez et al. in LMDB format
• af2_dark_durairaj_lmdb.zip: AF2 human proteome embeddings from Durairaj et al. in LMDB format

Run PARSE on a single PDB

Run the following to annotate a protein using standard CSA reference database.

From PDB file:

python predict.py --pdb PDB_PATH

From pre-computed embedding database, given a PDB id:

python predict.py --precomputed_id PDB_ID --precomputed_lmdb LMDB_PATH

Optional arguments:

--chain CHAIN: annotate only a specified chain
--db DB_PATH: reference database embeddings, in pickle format
--function_sets FN_PATH: reference database function sets, in pickle format
--background BKG_PATH: function-specific background distributions, in pickle format
--cutoff CUTOFF: FDR cutoff for reporting results (default 0.001)
--use_gpu : flag for running with GPU

Create embedding database from directory of PDB files

To run PARSE on a large number of PDB files, you can create a pre-computed embedding database in LMDB format for all pdb files in a directory (including subdirectories). Valid filetypes include pdb, pdb.gz, ent, ent.gz, cif

For small datasets, run without optional split arguments:

python embed_pdb_dataset.py PDB_DIR OUT_LMDB_DIR --filetype=pdb

For large datasets (e.g. Swissprot, AlphaFoldDB), we recommend processing in parallel using the num_splits argument.

python embed_pdb_dataset.py PDB_DIR OUT_LMDB_DIR --split_id=$i --num_splits=NUM_SPLITS --filetype=pdb

This produces NUM_SPLITS (e.g. 20) tmp files in OUT_LMDB_DIR. To combine all into the full dataset, run the following:

python -m atom3d.datasets.scripts.combine_lmdb OUT_LMDB_DIR/tmp_* OUT_LMDB_DIR/full

To run PARSE on each protein in the embedding database:

python run_parse_lmdb.py --dataset=LMDB_DIR

Optional arguments are the same as predict.py, with additional split arguments for parallel processing: --db DB_PATH: reference database embeddings, in pickle format --function_sets FN_PATH: reference database function sets, in pickle format --background BKG_PATH: function-specific background distributions, in pickle format --cutoff CUTOFF: FDR cutoff for reporting results (default 0.001) --split_id SPLIT: split id (int) representing index from 0 to NUM_SPLITS-1 --num_splits NUM_SPLITS: number of splits for parallel processing

Create new reference database

To create a new reference database of functional sites, first generate a csv file with the following columns (see data/csa_functional_sites.csv for example):

site: function ID from source database (e.g. M-CSA 993)
pdb: pdb and chain (e.g. 2j9hA)
locs: list of residue ids which are functionally important
source: database source (e.g. M-CSA; used in case of more than one source database)
description: text description of function (e.g. Glutathione S-transferase A)

Then, create embedding database using the following, where EMBEDDING_OUT is the generated .pkl file containing DB embeddings and FUNCSET_OUT is the generated .pkl file containing function sets:

python create_reference_database.py DATABASE.csv EMBEDDING_OUT FUNCSET_OUT