"Deterministic" Fast Relax

[nathankvphan]

Hi all,

I'm trying to evaluate the effects of mutations on structural stability. I'm currently using the Fast Relax protocol to do so, but I keep getting slight variations in my final values. My script has the '-constant seed' flag, so I'm not sure what is causing this variation.

I want to use Fast Relax because (to my knowledge), it gives me the most accurate results, but if I can't get this to work, are there other alternatives I should consider? I'd like to maintain backbone flexibility in my final script. Thank you!

[roccomoretti]

-constant_seed will only fix the random number generator at the very start of the program execution. If there's any change in how the program runs, you'll may come into the RNG at a different location, giving you different results.

For example, if you have multiple structures (or mutations) and run them in a different order, you'll get different results, even with a constant seed, as which RNG results you use for which structure will change. If you want consistency, you'll need to run things the exact same way each time. For command line Rosetta, that means providing the exact same options each time. For PyRosetta, that means doing the exact same things in the exact same order from the RNG initialization.

Note that with PyRosetta you can possibly re-initialize the PRNG. You should be able to call rosetta.basic.random.init_random_number_generators() to re-initialize the main PRNG. If you call this at a consistent location before each mutation (and use -constant_seed), you should get consistent results.

That said, the stochasticity in FastRelax is inherent. By fixing the seed you'll get consistent results, but they won't necessarily be better results. There's a distribution of results you'll likely to get out of the Relax, and each random run will give you a different one. There's no guarantee that -constant_seed will give you the best-scoring result, or even an "average" result. You could get unlucky and get an anomalously bad result for that particular structure/mutant. The results of relax are stochastic, and you have to account for that variability in your analysis. (e.g. by running the same relax multiple times, and looking at the spread, or simply realizing that you do get slight variation, and hopefully that variation is below the level that matters in most cases.).

Keep in mind that FastRelax and the Rosetta energy function is an approximation in the first place. Often the errors you get from Rosetta not being a 100% accurate reflection of reality are larger than the errors you get from the stochastic nature of the simulation. The more precise you need the results to be to see an effect in the calculations, the more careful you need to be in how you're doing the modeling -- versus simply needing the numbers the calculation spits out to be more precise.