Efficient protein-ligand interaction energy

[John Chodera]

Our application (yank) involves evaluating the potential energy of a protein + ligand system using OBC GBSA (no cutoff) at different alchemical states, where either (i) the interactions *between* the protein and ligand are attenuated until we have completely decoupled protein and ligand, or (ii) the ligand itself is annihilated by turning off its Lennard-Jones and GBSA parameters. We've chosen approach (ii) for now because we can't find any way to just modulate the GB interactions between protein an ligand. In principle, we only need the relative energy *differences* to these alchemical states from some reference state.

We're currently finding that energy evaluations are very slow, presumably because they are computed on the CPU. If only the ligand parameters are changing, and we only need energy differences, is there a faster route to computing these potential energy differences using the existing API? For example, for the NonbondedForce term, if we set the charge and well depth to zero, and added only protein-ligand interactions as Exceptions, would this be faster than computing all interactions by the normal mechanism? Is there a corresponding trick that could be used for the GBSA term?

[Peter Eastman]

You're correct that it's slow because the energy evaluations are done on the CPU. That should be much faster once we get the GPU energy kernels done.

I don't think there are any special tricks that will make it faster. Adding exceptions almost always requires more computational work, not less.

Peter

[John Chodera]

Hi Peter,

Thanks for your reply! We'll stick with computing the whole energy then, for simplicity.

I know there are a million things being done right now, but is there a rough ETA for energy evaluations on the GPU? This will help us decide whether it's worth the effort of hand-coding calculation of the protein-ligand interaction energies for the tests we're trying to run, or if we're better off just waiting for the GPU-accelerated energy routines.

John

[Kai Kohlhoff]

Hi John,

We are working on the potential energy kernels and have just completed those for bonded interactions. My time estimate for nonbonded interactions including OBC GBSA is roughly 2-3 weeks from now. They should be included in future releases of OpenMM following that date.

Kai