Md decoys

[ftclark3]

Description

The commit history is a little messy, but I think the end result is cleaner. I can clean up the history if you want.

In summary, I did the following:

• Add a setter method for the pdb_structure attribute of the AWSEM class that recalculates the pdb_structure-dependent attributes without reinitializing the entire class. This allows you to cleanly iterate over frames of a trajectory (as represented by Structures) and modify your AWSEM 's conformation
• Extract part of the AWSEM class into a new class, AWSEMBase
• Write a new class, AWSEMIndicators, that has all the essential attributes and methods of AWSEM but can be initialized from an array of indicator functions
• Write a new class, DecoyEnsemble, that computes statistics of the distributions of each indicator function over a set of conformers. Technically, this involves loading indicator functions for a set of frames as described in point (1), then computing the average or variance over frames. Each statistic is its own method, like DecoyEnsemble.average(self) and DecoyEnsemble.variance(self) , that runs only when called by the user. If necessary, I can have DecoyEnsemble.init write the indicator arrays in chunks to a file, and the average and variance methods can read them in chunks. The average/variance indicator arrays can then be passed into AWSEMEnergy just like the native conformer. The complete Wolynes foldability criterion would then be coded as (AWSEMEnergy(native) - AWSEMEnergy(average)) / AWSEMEnergy(std), which should only be about 3x more expensive than just optimizing the native energy, compared to a much bigger factor (don't remember exactly how big) for the current variance estimation

Todos

Notable points that this PR has either accomplished or will accomplish.

• Supports accurate and efficient MD decoy-based protein design as described above

Status

• Ready to go

[ftclark3]

This also opens up the possibility of md decoy-based frustration calculations

[ftclark3]

will revisit this soon