antoinepassemiers/gde-gaussfold

Lightning-Fast Template-free Protein Folding based on Predicted Residue Contacts and Secondary Structure

Graph-Distance Evolutionary (GDE) Gaussian Folding

GDE-GaussFold is a template-free protein modelling software that assign 3D coordinates to residues based on predicted contact maps and secondary structure. It uses Gaussian restraints to model distances between pairs of residues and searches for the model that maximizes log-likelihood. Optimization is done heuristically with a non-stationary genetic algorithm.

from gaussfold import GaussFold

# cmap is an array of shape (L, L) of
# predicted contact probabilities.
cmap = ...

# ssp is an array of shape (L,) representing
# 3-state secondary structure prediction.
# 0 stands for 'H', 1 for 'E' and 2 for 'C'.
ssp = ...

coords_predicted = GaussFold().run(cmap, ssp)

The library provides projection-invariant evaluation metrics for the predicted 3D models:

from gaussfold import tm_score, rmsd

# coords_target is an array of shape (L, 3)
# representing the 3D coordinates of the protein
# native structure. If at least one residue is unknown,
# then coords_target is a list of triples where
# missing residues are replaced by None.
coords_target = ...

print(tm_score(coords_predicted, coords_target))
print(rmsd(coords_predicted, coords_target))

A complete example is available in example/ folder.

Installation

GDE-GaussFold can be installed with the following command:

python setup.py install

Dependencies

• Numpy
• Scipy
• NetworkX
• Scikit-learn

Optional dependencies (but required in order to use the PDBParser):

• minineedle (https://github.com/scastlara/minineedle)
• miniseq (https://github.com/scastlara/miniseq)