Genetic Algorithm applied to Protein-Protein Docking
The Sparks Lab is no longer offering DCOMPLEX, which is a critical component of gdock (as its scoring function) - until this is solved or some other scoring function is implemented, gdock is broken 😭
Its easier to run via its server! Check it out at gdock.org Sorry, I had to take down the server. These things are not free, you know 😔
$ git clone https://github.com/rvhonorato/gdock.git
$ cd gdock
$ python setup.py develop
$ bash install.sh `pwd`For more informaion about the installation go here
$ cd examples/
$ gdock run.toml
For more information about the configuration run.toml file and how to setup your own runs go here
Despite having some knowledge in Genetic Algorithms I am not an expert. If you are an expert any suggestions or critics are very welcome. (:
In gdock we apply genetic algorithm to protein-protein docking. Here, each individual is represented by its cartesian coordinates and euler angles, and the fitness of each conformation is given by the ratio of restraints that are satisfied.
The input of each simulation is two PDB files, plus a list of residues that are likely to play a part in the interaction. There are many ways one can obtain such residues, from theorethical predictions to in vitro experiments. gdock does not support ab initio docking.
During spatial sampling, the receptor is fixed in space and all geometric opertions are be done to the ligand. Each individual (=conformation) is generated by randomly assigning three floats that will describe a region in space around the geometric center of the initial ligand positioning (in a range of ± 5Å) and three more floats that will describe its rotation (0-360).
The fitness of each one of these individual is evaluated as a ratio of how many of the inputted restraints are observed to be in contact (distance < 4.9Å). Over each generation, individuals have a chance of mutation, where one of the six descriptors are randomly changed and also a chance of crossover, where individuals exchange descriptors.
The generations are stopped after it "converges" (mean variation > 0.1 over 3 generations) and the binding energy of all conformations are calculated with DComplex. The conformations are ranked according to the gdock_score which is given simply by:
gdock_score = energy / restraint_satisfactionThe top 1000 structures are clustered using according to their frequency of common contacts with cutoff=0.6.
Once the simulation is done, the run directory is cleaned to preserve disk space and a data-processing friendly file is written in analysis/gdock.dat
The accuracy of a docking software can be measured by its ability to reconstruct experimentally determined structures. The dataset most used for such benchmarking is the Protein-Protein Docking Benchmark v5 (BM5) (10.1016/j.jmb.2015.07.016).
A gdock run was done for each of the complexes of BM5, using the unbound forms of the ligand and the receptor as input, and calculating the interface root mean square (i-RMSD) variation of the resulting complexes against the bound form. The restraints used in the benchmark are the true-interface, defined as the aminoacids that are observed in the bound interface. These restraints can be interpreted as the perfect scenario, thus ideal to measure gdock perfomance.
According to the Critical Assessment of PRediction of Interactions (CAPRI) parameters, comlexes can be categorized as High, Medium or Acceptable according to its (i-RMSD):
i-RMSD <= 1Å - High
i-RMSD <= 2Å - Medium
i-RMSD <= 4Å - Acceptable
Single-structue: For each target in the benchmark, we could rank the generated complexes according to its score and, for example, analyse only the top 5 conformations. If at least one of these conformations has i-RMSD <= 2Å, this target's success will be considered Medium.
Cluster-based: Once the clusters are sorted by fitness and we consider, for example, the top5 models of the top 1, 3, 5 or 10 clusters. Same logic as the single-structure analysis is applied.
The results for v1.1.0 are as follows, the % shows the increase/decrease in relation the previous version.
Note: Some conformations were excluded since dcomplex could not be executed on them.
Unfortunately, no High ranking conformations were obtained and there was a decrease in success rate for top 1, 5 and 10. Even tho the exact reasons for this behaviour are not certain, there are a few possibilities: 1) the scoring function is not perfoming well at positioning the "best" conformations in the top rankings, 2) the range of spatial exploration might be too large or 3) the lack of flexibility during docking is affecting the success rate.
A notable increase (300%!) in success rates is observed for the top 200, 400 and 1000 both in Medium and Acceptable ranking, when comparing the benchmark results with the previous version. This comes mostly due to the inclusion of the restraint satisfaction as a score term however these results indicate that the scoring function can still be improved.
As a better scientists said before me:
Scoring is the holy grail of docking!
Note: All relevant benchmarking data are kept in a different repository.
If you would like to help or improve your python coding skills, gdock might be a good project to work on!
The code is licensed under a 0-clause license (0BSD), which pretty much means you can do whatever you want with it; copy, edit, deploy, distribute or even sell it to the highest bidder (good luck with that).
As for my vision to the project, there's a small TODO list.
🐙
