Adaptive free energy calculations (with refactored API)

[hannahbrucemacdonald]

Trying to enumerate all the ways that perses can be used, and might be used in future

[hannahbrucemacdonald]

• [ ] relative ligand-protein calculations

• static networks: either star-shaped, lomap, fully-connected or a user-defined connectivity

• dynamic networks: adaptive sampling, with weights on edges

• [ ] relative protein mutations

• single point mutations

• multi-point mutations

• [ ] protein:protein mutations (i.e. the mutation might happen on either protein)

• [ ] ligand and protein mutations (i.e. a set of ligands and protein mutants)

• [ ] multi-phase mutations, where a small molecule or protein is mutated in two of:

• vacuum

• solvent (water) (relative hydration)

• solvent (octanol or other) (relative logP)

• membrane (partitioning)

• amorphous solid for solubility

• different protein complexes

[hannahbrucemacdonald]

In future, it might be useful to have networks that are run in combination with other packages -- i.e. YANK or ROCSALT

[zhang-ivy]

Looks good -- the only thing I can think of that's missing is in the last item, where we also want to support functionality for mutating an apo protein across phases

[dominicrufa]

the only way i see that we can allow support for all of this functionality is to make a network where every (non-alchemical) node is indexed by a (environment, complex) tuple where environment tells us what our system is solvated in and complex tells us the what our system actually is (protein-ligand/protein protein, just ligand, etc). after this network is specified, there has to be a class that can handle (at a high level) figuring out what alchemical transformation edges need to be built so that the appropriate relative binding free energy can be computed.

[jchodera]

The use cases look great!

In future, it might be useful to have networks that are run in combination with other packages -- i.e. YANK or ROCSALT

We may instead just end up integrating all of those capabilities into perses!

@dominicrufa : When @jaimergp and I worked this out on the whiteboard, we thought each edge would connect two (protein, ligand, environment) pairs that share the same environment. You can think of this as having parallel graphs where each layer is a different environment, and each physical quantity corresponds to differences in free energies between different pairs of nodes in different environments.

[jchodera]

The design notes from @jaimergp and I can be found on this gist.

[hannahbrucemacdonald]

• amorphous solid for solubility calculations