addMissingAtoms additional params

[scschneider44]

Would it be possible to add ways to tune the internals of addMissingAtoms?

Specifically it would be nice to be able to pass maxIterations to the calls to the LocalEnergyMinimizer minimize calls (https://github.com/openmm/pdbfixer/blob/master/pdbfixer/pdbfixer.py#L1291 and https://github.com/openmm/pdbfixer/blob/master/pdbfixer/pdbfixer.py#L1321) as well as to tune the number of steps here https://github.com/openmm/pdbfixer/blob/master/pdbfixer/pdbfixer.py#L1312.

This would allow us to trade off accuracy and latency for large proteins. Right now some proteins are prohibitively slow to fix given the size of server we're using.

[peastman]

Yes, it would be possible. Do you have an example where it's unnecessarily slow? I'd like to see where the time is being spent.

[scschneider44]

Thanks for looking into this.
7E2Y is slow (>30 mins) for me on a machine with 4 vCPUs. The lines in the first message were the functions taking the most time for me but I didn't drill down beyond that.

[peastman]

I just tested it. On the first screen I told it to download 7E2Y from RCSB. On the second screen I told it to keep the four protein chains and delete the final chain that contains heterogens. On the third screen I kept everything selected and clicked continue. It took only about 10 seconds to add everything. That's on my laptop with a low end embedded GPU!

What kind of hardware are you running on?

I don't think you actually want to add all the missing residues though. In one case it's adding a stretch of over 150 residues to the end of a chain, and in two others it's adding more than 100. It can't do anything sensible with them.

[scschneider44]

Sorry I missed this. We are kind of running on a potato in digital ocean. It has 4vCPUs no GPU and 8GB RAM.

Is there anyway to tune the method so that a stretch like that isn't added?

[peastman]

What kind of CPUs are they? What architecture, what generation, how many cores?

[scschneider44]

They're intel CPUs (so x86). Digital ocean doesn't expose the exact CPU but say "first generation or older Xeon Scalable processors".

A vCPU is ~= 1 core. Per digital ocean:
"A vCPU is a unit of processing power corresponding to a single hyper-thread on a processor core. A modern, multi-core processor has several vCPUs."

[peastman]

So basically it's equivalent to a 2 core CPU, since Xeon processors have two hyperthreads per core. Also, it's ancient. The first generation Xeon Scalable was introduced in 2017, and they say it's that or older. That may just not be enough for what you want to do.

[scschneider44]

Okay thanks for looking into it