Preview: New confidence constraints for RibonanzaNet-SS

Hey all! You may remember that when we introduced RibonanzaNet-SS, we also added metrics to the specbox called eF1 and eF1,cross-pair which were presented in the paper as a way to determine how accurate we predict the model’s prediction is. However, there were a number of limitations in how that metric was formulated, and over the past few weeks we’ve been working on some alternative metrics that we hope will be more accurate and reduce “weird” behavior. We’ve also implemented them as constraints to make it easier to refer to

A demo of how this will work is available on our development site using a variant of OpenKnot Round 4 Week 1 (SARS-CoV-2 frameshift 6XRZ). The graphics are temporary, thrown together by me while Sharif (our designer) comes up with something better :slight_smile:

Looking for a couple bits of feedback:

  1. Does this score appear usable/“solvable”?
  2. Any input on how we are displaying these as constraints?

We hope to make this available in the OpenKnot Round 6 labs early this week as folks work on their final submissions to the lab before it closes next weekend. Depending on how things play out, we may make this available more broadly in the specbox - or make further adjustments.

For those interested, some implementation details:

  • These metrics work similarly to the existing “TEA” (Target Expected Accuracy) metric in the specbox. However, instead of using the MCC (Matthews Correlation Coefficient) formula, we use the F1 formula. The idea is that instead of a classic F1 where you characterize a prediction as “positive” or “negative” (eg, whether a pair was predicted or not) and “true” or “false” (ie, whether the prediction is accurate or not - which we don’t know yet!), we use weighted values for these categories based on the base pairing probabilities.
  • We have variants of these constraints both with respect to the target structure (ie, how confident are we that the sequence will fold into the target) as well as with respect to the natural structure (ie, how confident are we that the sequence will fold into whatever the model predicted).
  • For the pseudoknot variants, what we do is “zero out” the contributions of any bases that are not predicted to be “cross pairs”. This means unlike the “global” version, we don’t necessarily include information about whether there are missing cross pairs - instead we are focusing on whether the cross pairs that are predicted are correct.

The higher each number is the better is the assumption? I would think, however, that designs should not be abandoned just because these numbers do poorly. Correct?

1 Like

In the case of natural confidence: This is roughly the likelihood that RibonanzaNet-SS’s prediction is correct. So it’s “better” in as much as if you like RibonanzaNet’s prediction, it’s more likely that’s what’s actually going to happen. Lower —> RibonanzaNet-SS doesn’t really know what the right structure is.

In the case of target confidence: Higher —> RibonanzaNet-SS thinks there’s a greater chance the target is what’s going to happen. Lower —> RibonanzaNet-SS doesn’t think the target structure is likely to happen (though with a low natural confidence, it more likely may not know what’s actually going on at all!)

At least, that’s the general idea. The normal caveats still apply as always that the model could be wrong even if it thinks it’s right :slight_smile:

1 Like

Hi @LFP6

I have been playing around with the test version.

I’m trying to understand why the numbers in two first boxes are identical to the numbers in the two last boxes:

0.967, 0,981, 0,967, 0.981

CUCGUUUUUUCGUUGAUGCAGUGCGUACGGCUUGCGACGCACUGCGCUAUACUGCCUGGAGACUGCAUGUAUAGAAUACCAAGCACAU

Is it because native and target are even?

Box design

On the images inside the boxes, it would be really helpful if the numbers are much larger, since that is the main info.

Similar I like the target and leaf symbol to be larger. If the numbers could symbolize by the red, yellow or green colors where on the range they were, that would be more helpful than the similar graph images.

I miss the cross-pair and ef1 naming.

Model improvement

I particularly like that it now seems to be giving a better bonus to actually sticking GU’s into longer stems. The current version in lab tolerates it but most often don’t give it extra bonus. I only found GU’s needed for upping the RNNet score in very few of the 17 weekly puzzles. Most of them I could solve with 100% GC

Whereas one of our earlier versions was tolerant to GU’s only at ends of stems. (Think it was the mini version made over PDB pseudoknots) So it looks like things are moving in the right direction.

I have been making naughty designs on purpose, to flag them to the algorithm. While still trying to keep the RNNet score high.

Astromons r6 list - Does natural match target?

Afterthought

Seeing it was possible to further improve my earlier cross pair topscore score in the newest model, I think there is more work to be done on maxing cross-pair score, amount of GC’s, AU’s and GU’s. I bet some of you can do a better job than me.

Can we get this model integrated in all the 17 subpuzzles now? While the graphics are temporary, I think it is helpful. Plus we are already digging into our new slots. The sooner we get the integration, the faster we can put this model to the test, by getting lab data back on it.

Yep, exactly!

1 Like

New constraints are now available in OpenKnot Round 6!

1 Like

Thx! :slight_smile:

By the way I noticed the new RNNet has a really cute bias. Probably because it seemingly flies in lab too.

RNNet model:

Lab result:


https://eternagame.org/game/browse/13259035/?filter1=Id&filter1_arg1=13278175&filter1_arg2=13278175

Off cause by the doing of Ucad. :wink:

Many of my cross-pair maxing designs ends up with a lot of U’s in the single bases. It is quite happy about lots of A’s on the single bases too.

I suspect though that the model is a bit too happy about the single U’s and A’s though. But that isn’t a bias of the new model alone, but also the earlier RNNet models we have had.

I have made A and U maxing series to flag it to the model.

When I look at the read count for Round 1 data, 6+ poly U looks like they are bad news. But the designs themselves do get a decent score.


Eli’s new Copy of Discoveries Round 1

When I scroll through puzzles it looks like the new constraint boxes do not update but default to the first one.

2 Likes

Yep, I see the issue - working on a fix

Sorry for the delay - this is now fixed!