README.md

RLFold

Fold 2D HP chains using reinforcement learning. A work in progress.

Dependencies

Currently depends on HPSandbox.

Moves

This implementation is currently limited to two-dimensional chains. The size of the grid may be set by the user.

A residue in a chain may execute one of the following moves: (i) an end-move or (ii) crank-shaft.

Objective

Find a policy $\pi^*$ that folds an HP chain into a ground state in as few moves as possible. The ground state(s) of a chain having some sequence are the configurations having the maximum possible number of hydrogen bonds.

Options

Consider a simple HP chain with sequence HHPHH.

Generate a configuration for the chain (relies on HPSandbox) --

cd examples
sh make_config.sh
cd .. 

Next, sample all configurations that are accessible via the move set

python config_generator.py --chain_length 5 --grid_size 7 --conf_dir examples --draw True

which should print out 13. A file containing the configurations is saved in the default directory, here /examples/5, along with configuration images and a movie.

Finally, compute the optimal policy using value iteration

python iteration.py -s HHPHH --draw True --verbose True