A Markov process preserves detailed balance if the ratio transition probability between any two states states is equal to the ratio of the probability of the states themselves. If it can be shown that it does, then the Markov process is reversible and evolving the system will equilibrate it to the stationary distribution.
The detailed_balance scientific benchmark tests if the BackrubMover, SidechainMover, and SmallMover satisfy detailed balance.
The first test defines a Markov process where the states are the conformations of an 8 residue poly-alanine peptide and the transitions are those defined by the BackrubMover with weight 0.99 and the SmallMover with weight 0.01. The process is evolved for 5,000,000 steps.
The BackrubMover satisfies detailed balance if the observed distribution over all degrees of freedom is uniform according to the spherical coordinate Jacobian (i.e. uniform for azimuth angles and sinusoidal for zenith angles). To test this, the DOFHistogramRecorder is used to construct a histogram over the torsion angles (PHI) and bond angles (THETA) in the backbone. The detailed_balance scientific benchmark records the mean squared error (MSE) away from the expected uniform distribution. Given the amount of sampling used in the test, the expected average MSE should be below 1e-5. Here is an example of a successfull run:
BackrubMover average PHI MSE : 6.10506e-06 BackrubMover average THETA MSE : 1.98768e-07
Here is an example of a successfull run:
SmallMover average MSE : 3.01119e-06
The second test measures the detailed balance of the SidechainMover, which samples off-rotamer sidechain conformations. By default, SidechainMover samples sidechain conformations with respect to Boltzmann distribution associated with the dunbrack rotamer energy. When testing the detailed balance, the SidechainMover samples sidechain conformations uniformly over the torsional angle degrees of freedom.
For each of the 20 canonical amino acids, a single residue pose is constructed and sidechain conformations are sampled with a within rotamer weight of 0.45 for 200000 steps. The detailed_balance scientific benchmark records the mean squared error (MSE) away from the expected uniform distribution averaged over all amino acid types. Given the amount of sampling used in the test, the expected average MSE should be below 1e-5. Here is an example of a successfull run:
SidechainMover average MSE : 9.23451e-06