Prefered Folding Engine for Labs? (Single-input switches, revisited)

Is there a prefered folding engine for the new lab? Seems like most of the submissions are from Vienna, but isn’t NuPACK more accurate? (I just found that there is a choice)


Not sure what makes you think that NuPACK is more accurate. It is the only choice we have when we’re dealing with 3 or more strands, so that’s what we were forced to use for OpenTB, but that is not a statement about accuracy or anything.

In fact, many experienced players largely prefer Vienna2, even though that doesn’t mean that this engine is any better. All our folding engines have strengths and weaknesses. I documented some of my own personal in-silico experimentations in a couple blog entries:

I also touched on some related issues during EternaCon 2016 (see chapter 4)

And you can find an example in my other current thread where Vienna gets somewhat closer to the correct folding of the Arginine aptamer than NuPACK does.

I’ll end up here with my usual disclaimer: I’m no scientist in this field, so please make sure to stay skeptical about anything I say on this topic.

1 Like

Ty. I was assuming that NuPACK was more accurate because it is so much slower than both Viennas. Seems like the usual trade off when creating an algorithm. Of course I don’t know how well these are optimized and the hardwere they are running on. And that’s why I asked.
Well, if it doesn’t really matter I’ll try to use them all equally and see the result :slight_smile:
Maybe one more question… Is it better to search for solutions that work for all three engines, or would that be pointless?

If I recall correctly, at least one of the devs (rhiju) has suggested multiple times that it could be interesting to find sequences that achieve consensus over all the engines that we have. I’m skeptical about that, but it’s possibly worth the try. Unfortunately, this may prove quite tough to achieve (as you saw, there are some pretty big differences between engines).

Good luck :slight_smile:

(oh, and NuPACK’s “slowness” is just an artifact of the quite unoptimized source code, Vienna is comparatively highly optimized, a bit too much so in my opinion)

Quite tough indeed, but it seems possible. It is relatively easy to fulfill the lab criteria across all three engines, but they usually show different structures. However even that can be adjusted after some tries. Here is my “Exclusion - Tryptophan A” attempt:

Picture me impressed (not being sarcastic). Well done :slight_smile:

1 Like

:slight_smile: This is actually a fun challenge. Found one for “Same State - Tryptophan B” as well:

Vienna is the default. Is it wrong for me to assume since its default its probably the best choice? my main question is what engine is used for the actual test in the real lab? Or if an engine isnt used in the actual lab what engine is more similar to it?  Thanks

In the actual experiments, the folding engine is called Nature. And Vienna is neither the best choice, nor the default, it simply used to be the only engine we could use for puzzles and that’s why a large portion of the collection of player puzzles is using that one. When working in recent labs, and if the targets allow for it (exceptions are for instance puzzles that require NuPACK because they have 3+ strands), you can use any folding engine and the applet will remember which one you used last.

Yes to consensus solves, especially those that have identical or nearly identical secondary structures! This could help build better models too. I wonder how they will synthesize.

Yes to consensus solves, especially those that have identical or nearly identical secondary structures! This could help build better models too. I wonder how they will synthesize.

Like Pi above , I have noticed and now strive for consensus solves of all three folding engines and their secondary structure and even their dotplots  (as much as possible). It can be done. It should be a useful test for improving models and correlating to actual synthesis results.

Here is one example, much like Pi’s above:

Probably too late but if you put “_#3_” for “works for all 3” somewhere in your description it makes seeing if the strategy works much easier. For mine: “_Nu_” means Nupack and Vienna otherwise.

If, as the name suggests, Vienna2 is a new, improved version of Vienna (washes whiter!) then is the original Vienna really needed?

With each model having their own strengths and weaknesses a safe bet would to be to have designs that pass all three engines I’m assuming

I kept track of my submissions that passed all 3 a few labs ago and it didn’t seem to make a difference, a lousy design is a lousy design. My test wasn’t all that scientific however. I’ve
been using Nupack for this lab.I am interested in what you come up with and hope your are able to arrive at a different conclusion.