AlphaSync (https://alphasync.stjude.org) is an updated AlphaFold structure database synchronised with UniProt.
AlphaSync predicts new structures to stay up-to-date with the latest UniProt release, and it additionally enhances all structures with residue-level data such as solvent accessibility and atom-level non-covalent contacts. AlphaSync is described in Lang et al., Nat Struct Mol Biol 32, 2628–2632 (2025), https://www.nature.com/articles/s41594-025-01719-x.
This repository provides the structure prediction and processing pipeline behind AlphaSync. It is not yet optimised for local deployment. It currently requires a local SQL server to be set up and is optimised for an LSF HPC environment with a local Singularity container of AlphaFold 2.3.2. We have tentative plans for a more portable containerised version in future.
- Requires Lahuta for residue-residue contacts, which is not yet publicly released (but should be soon) (https://bisejdiu.github.io/lahuta)
- Install Python packages:
pip install biopython gemmi ijson ipdb natsort numpy pandas pymysql requests sqlalchemy tabulate tqdm- Install the Google Cloud CLI so that gsutil is available (https://cloud.google.com/sdk/docs/install)
- Install DSSP to make sure mkdssp is available (https://github.com/PDB-REDO/dssp)
- Update blang_mysql.py with SQL connection details
- Create tables in sql/sql_create_statements.sql and import .sql files
- Run
python run.py- Downloads structures from AlphaFold Protein Structure Database (AFDB) via FTP and GCS
- Parses structures
- Calculates RSA/dihedrals/contacts
- Maps sequences to structures
- Run
python run.py -alphasync(add-refreshto run even if the local UniProt release is still current)- Refreshes protein sequence and proteome data from UniProt REST API and FTP
- Submits AlphaFold structure prediction jobs as needed
- Parses structures
- Calculates RSA/dihedrals/contacts
- Maps sequences to structures
- Can then migrate alphasync_compact SQL tables to web server (code available on request)
- Repeat for new UniProt releases
The code in input/alphasync/alphafold_tools is modified slightly from https://github.com/google-deepmind/alphafold, licensed under the Apache 2.0 license. The main change is a split into CPU- and GPU-based steps for more efficient parallelisation, similar to AlphaFold 3.