Graph-based community clustering approach to extract protein domains from a predicted aligned error matrix

Using a predicted aligned error matrix corresponding to an AlphaFold2 model (e.g. as downloaded from https://alphafold.ebi.ac.uk/), returns a series of lists of residue indices, where each list corresponds to a set of residues clustering together into a pseudo-rigid domain. Two different implementations are provided, using either NetworkX or iGraph (default) as the back-end. Results in either case appear broadly comparable; being written in compiled C code iGraph is significantly faster (>40X in core routine, about 5X faster overall runtime).

• Python >=3.7
• NetworkX >= 2.6.2 OR python-igraph >= 0.9.6

Due to an internal implementation issue in NetworkX 2.6.2 (Issue #4992) some combinations of PAE matrix and resolution can lead to a KeyError. This will be fixed in the next NetworkX release.

While primarily intended as a code snippet to be incorporated into larger projects, this can also be called from the command line. At its simplest:

python pae_to_domains.py pae_file.json

... will yield a .csv file with each line providing the indices for one residue cluster. Full help for the command-line version:

positional arguments:
  pae_file              Name of the PAE JSON file.

optional arguments:
  -h, --help            show this help message and exit
  --output_file OUTPUT_FILE
                        Name of output file (comma-delimited text format.
                        Default: clusters.csv
  --pae_power PAE_POWER
                        Graph edges will be weighted as 1/pae**pae_power.
                        Default: 1.0
  --pae_cutoff PAE_CUTOFF
                        Graph edges will only be created for residue pairs
                        with pae<pae_cutoff. Default: 5.0
  --resolution RESOLUTION
                        Higher values lead to stricter (i.e. smaller)
                        clusters. Default: 1.0
  --library LIBRARY     Graph library to use. "igraph" is about 40 times faster; "networkx" is pure Python. Default:
                        igraph

Using https://alphafold.ebi.ac.uk/entry/Q9HBA0 as an example case...

resolution=0.5:

resolution=1.0:

resolution=2.0: