[A]/[B] discussion

  • 3
  • Article
  • Updated 3 years ago
  • (Edited)
This is a new thread for analyzing [A]/[B] puzzles, starting with the puzzles in R100
We are very excited about the results which have been posted in Eterna:

and the data can be found in
Google spreadsheets: Eterna R100 [A]/[B]
or the Excel file: Eterna R100 [A]/[B]
Photo of johana

johana, Researcher

  • 96 Posts
  • 45 Reply Likes

Posted 4 years ago

  • 3
Photo of Omei Turnbull

Omei Turnbull, Player Developer

  • 977 Posts
  • 305 Reply Likes
Great news, Johan!

Some of the (sub)score columns look funny.  If I sort by Eterna score, here's the top of the list:

The first line has the normal subscores for an Eterna score of 100, but the following lines don't.  Is this on purpose?
Photo of johana

johana, Researcher

  • 96 Posts
  • 45 Reply Likes
Yes. The explanation will be put online very soon. The reason is that the MS2 control puzzles only have one state, so they are not switching and have no switch score. Also, the MS2 control OFF is actually good is if no binding is seen, so there is also no folding score in the normal sense. The other puzzles are scored as before but the most illuminating variable is often the fold change.
Photo of Omei Turnbull

Omei Turnbull, Player Developer

  • 977 Posts
  • 305 Reply Likes
Thanks.  That makes perfect sense now.  Look forward to your commentary and graphs.
Photo of Eli Fisker

Eli Fisker

  • 2229 Posts
  • 490 Reply Likes
I have uploaded the data in the google sheet to a google fusion table:

Eterna R100 [A]/[B]
Photo of Omei Turnbull

Omei Turnbull, Player Developer

  • 977 Posts
  • 305 Reply Likes
... and here is a version with switch graphs added.
Photo of Eli Fisker

Eli Fisker

  • 2229 Posts
  • 490 Reply Likes
Couldn't help but look at the data. :)

R2 (2 State model)

The R2 lab winners made by JR showed the overall same tendency.

  • Sequence B came before A

  • The MS2 was sharing lane with sequence B

  • MS2 in between the oligos

JR design, Score 97%


The general trend was having the MS2 in between the oligo complements and using the main part of the design up. This is similar to what we have seen in the logic gate labs already.

R2 Drawings overview

R2 Majority2jpg

R2 Minority2jpg

The more unusual bit in the majority solves, are that the B sequence were doing a crossover pairing up (which is more usual in microRNA turnoff labs) in the majority of the winners. With single inputs that is.

Majority type

R3 (3 State model)

Here the jandersonlee winners had extensive sequence sharing between both sequence A and B but also B and MS2.

  • A sequence before B

  • A and B sharing lanes - or really magnet segment :)

  • B and MS2 sharing lanes

  • MS2 on the outside - not between oligo complements

jandersonlee winner, score 99%


Here the main trend was using only part of the design sequence and having the rest of the excess bases tucked away as static stems.

R3 drawings overview

R3 Majority2jpg

Perspective on R2 versus R3

The R3 winners follow a true blooded exclusion solving style. By the choices made, they make sure that the unwanted sequence/MS2 is either included or excluded.

When I say exclusion solving style I mean a solving strategy where the switch is made by sequences sharing lanes and where one will actively exclude the other by binding and vice versa.

The exclusion style as followed in R3, is a short cut way to achieve a switch. It is the solving style that is simplest, requires least space and involves fewest switching pairs.

This strategy has been extremely successful in the microRNA-in, reporter-out labs already. However these labs did not have to have the MS2 to fit in the equation too. Which I think is what complicates things, when comparing the reporter labs with the R2 and R3 labs.

The exclusion style takes extensive sequence sharing. And from both sides of the sequences - no matter which order the oligos are placed in relation to each other. Plus the same goes for sequence sharing in relation to MS2 too. Which again takes all the sequences being a match for this in the first place.

I did make an exclusion style series of designs in R2, where A and B shared lane, but none of them scored well. (With a max of 65%)

So if the overlaps between sequences are not good enough - then the best road to get to a solve, seem to be the one that involves going more complex and using a far bigger part of the RNA sequence to solve the problem. By placing the MS2 a middle spot and to place some of the switch elements further apart, to give space to put sequences in between for doing the switch job. To either turn on or off MS2 or sequence inputs.

Instead of having the switch elements themselves do the job directly on each other. As is done with pure exclusion style.

The R2 winners do something kind of in between exclusion style and having sequences in between the switch elements do the job. They still use the fact that MS2 and sequence B can be made overlap and share sequence. And as such get half the benefit from the exclusion style, plus escapes its downfall, due to sequences not matching well enough for a successful overlap, when the oligos are reversed. R2 is oligo B comes before A where R3 is oligo A comes before B.

The morale of this story is that it is easier to get two sequences (switch elements) to share 1 overlap nicely for taking turns, than to make 3 sequences (switch elements)  share 2 overlaps evenly between them and take turns. The R3 results show that the latter is possible, but the R2 results underlines that it seems to be harder. 

It seems - no surprise - that the exclusion style strategy is absolutely strongest - but only when the sequences are able to overlap perfectly with each other for achieving exclusion.

Photo of rhiju

rhiju, Researcher

  • 403 Posts
  • 123 Reply Likes
Eli – many many thanks for this initial distillation!

Are these conclusions based on looking at one or two winners (and perhaps their mods)? Or do we have reason to believe that these are statistically significant observations?

Getting diverse solutions is probably going to be most important for the project if we want to learn rules that generalize to other A's, B's, and indeed to A*B/C^2. That's the expression we will want to compute soon in the openTB challenge.

For example, the rule of A before B in the 3-state (R3) -- is that present in a whole bunch (>5) of diverse successful solutions? If its just present in one or two (+mods) perhaps this is a statistical accident – the fact that the R2 top solutions have different 'rules' but are supposed to have the same function suggests that to me.  

If we do not yet have a lot of diversity, perhaps our next step will be to challenge players to try hard to find alternative topologies in the current lab round, which repeats A/B! Rather then try to find mods of the current successes, which is my guess for what will happen. 
Photo of Eli Fisker

Eli Fisker

  • 2229 Posts
  • 490 Reply Likes
Hi Rhiju!

Np. I will try address your questions below.

Questions and answers

Rhiju: Are these conclusions based on looking at one or two winners (and perhaps their mods)?

Or do we have reason to believe that these are statistically significant observations?

Yes, in some cases my conclusions for the individual lab are based on winners around the same mod and thus following same structure pattern.

Typically I’m interested in general patterns. Both among winners and high scorers. This is why I try call out a majority type patterns - indicating that there are general more good designs following this pattern. But I also call out minority patterns, like if there are a group of designs following a different pattern, but either not seem as good as the majority pattern designs or there simply are a lot fewer of them.

The "A before B" rule

Rhiju: For example, the rule of A before B in the 3-state (R3) -- is that present in a whole bunch (>5) of diverse successful solutions?

For the R3 lab, there is only 3 winners. All the designs that score in top are of this type. This isn’t a lot.


However I’m not basing the A before B rule, or rather the leaving sequence gets before the staying sequence, on just this lab. This pattern pops up in other labs too.  

This even happens in the miRNA-in, reporter-out labs, round 1 and 2 that don’t have a MS2. These two images each represent the majority group among the winners and top scorers.

miRNA-in, reporter-out, round 1, score 100% by Omei


miRNA-in, reporter-out, round 2, score 100% by Jieux


More analysis from that lab

Note that A and B can be different sequences. It could just as well be B before A. It just happened to be an A before B in that particular R3 lab.

The sequence that needs to leave (stay in first state and leave in 2 state) - is before the sequence that needs to stay (be absent in state 1 but stay in second state).

Thinking about it, I should specify that this particular rule, seems most fitted to when the sequences actually overlap.

Reversed sequence pattern

For non overlapping sequences it is very well possible also having the opposite pattern. It can be reversed as was shown in the predefined binding site (alternative) lab.


Also notice that this design has far more heavy regions pair up that is normally needed.

Notice that here the B sequence which has to be kicked out, is placed last and the weak A sequence to be kept is placed first. However these sequences were far apart and not overlapping as was the case in the miRNA-in, reporter-out ones or the R3 solves. For more analysis on the rest of the A/B labs.

Sequence order is also affected by such things as sequence strength besides their position. Plus the balance between them can be shifted in multiple other ways, to allow for a reversal. Eg by oligo complement prolongation and magnet tails. We will actively need sequence reversal for some of the logic gate labs. This can also be seen in my expectations for the coming lab results.

Rhiju: If it's just present in one or two (+mods) perhaps this is a statistical accident – the fact that the R2 top solutions have different 'rules' but are supposed to have the same function suggests that to me.  

Actually the R2 winners has the leaving sequence (in state 1) before the staying sequence (in state 2). Just as R3.

The real difference between R2 and R3 is that R3 manages to have the MS2 overlap with both sequence A and B. Where R2 only manages to have its MS2 overlap with sequence B.

I think sequence sharing makes for a shortcut for making a working switch. However it will only work if the sequences are “chosen” so they are perfect for making a good sharing fit. We got lucky for R3 and seemingly not for R2 - at least not yet.

As I mentioned earlier, I made a R2 series, where A and B shared lane, but none of them scored well. (With a max of 65%) Perhaps some may get lucky getting my R2 design series a better score next round? It might work if the sequences are slided a bit in relation to each other.

As for the real tuberculosis lab, this sequence overlapping party where every sequence is overlapping with each other strategy will likely not work. At least not if used solely, as we need to have many sequences interact. Sometimes the bases added in between the oligo complements are the oil that make the switch glide.

However I have been saying that reusing structures between states seems to be beneficial. Having the sequences share lane and overlap is the ultimate sequence/structure reuse and was what happened in the R3 lab.  

Rhiju: If we do not yet have a lot of diversity, perhaps our next step will be to challenge players to try hard to find alternative topologies in the current lab round, which repeats A/B! Rather then try to find mods of the current successes, which is my guess for what will happen.

I would count both approaches valuable. We learn a lot from testing on the designs that works - just seeing that the few winners we got in first rounds, really were not by accident is worth a lot. Plus it is far easier doing experiments and break stuff on purpose in designs if we have an idea that they are good already. But I will like to see more of the latter approach too.

Placement bias in single input labs

Sequence placement bias also happens in the single input labs.

For the single input labs, the rule of thumb seems to be that the sequence complement and the MS2 should be put at whatever side, that keeps the MS2 away from the end of the design. Something I also gets into here.

Similar, if it is an ON lab, the MS2 should be before the sequence binding site. If it is an OFF lab, the sequence binding site should be before the MS2. Although the pattern is a bit more messy here. Sometimes there can be a sequence split as in the microRNA labs Sensor variant 1 and 2.

MS2 ON OFF overviewjpg

If the sequence is real strong, as is the case with the B sequence, a lot more crazy solving styles that stray from the above mentioned will be acceptable.

Wondering and bewilderment

This all started with me seeing RNA show a particular behavior for the A/B labs.

Shortly put: The late bit in the RNA sequence (3’ end) is the best landing spot for whatever one wants turned on. The early bit in the RNA sequence (5’ end) is best landing spot for whatever one wants turned off. In particular when it comes to single input labs.

I have been wondering about what is causing this sequence placement bias to happen. What is it about one end of the RNA that makes it a better spot for landing or a better place for kick off?

Is the one end of RNA stiffer than the other? Due to stacking/RNA’s chiral nature?

I have also been wondering how big a role MS2 play for all of this. If a RNA sequence of a certain length, was to be made pair up with a single sequence input (No MS2) and the binding site was placed at different spots along the RNA design. Would this discrimination difference show up too, like the 3’ end still being the best spot for pairing up?

I wonder about one more thing. If the labs were turnoff labs, would the placement trend be the same for double input labs? Seemingly the ON and OFF of the MS2, for single input labs are enough to change the favorable binding spot. Although to a lesser degree for the labs with the strong B sequence. This one allowed more odd and varied binding sites.

Photo of Eli Fisker

Eli Fisker

  • 2229 Posts
  • 490 Reply Likes
Pattern overview

I have made a pattern sum up for the single input labs and the labs with predefined sequence placement. I have been interested in what main patterns turn up among the winners and top scoring designs in these labs.

Basically this is a hint for what I think will work for next round.

Sensor A MS2 off

The top scoring designs follows this pattern.
  • A before MS2

  • A at bad spot (5’ side) - easier for kickoff

  • Design complex at 5’ end

Inspired by Omei

Sensor A MS2 offjpg


Sensor A MS2 ON

  • MS2 before A

  • A sequence at best turn on spot at 3’ side

  • Design complex at 3’ side

These trends are overall followed for all the winners.

Score 100, design by Brourd

Sensor A MS2 onjpg


The designs that do absolutely best for fold change, keeps a magnet G dangle at the tail. This do even better than a similar magnet G dangle at gap position.

My designs scoring 100% also keeps this magnet dangle, but have a different MS2 turnoff. One I transported from the jandersonlee’s winner in the microRNA 208a lab. Was real nice to see it work here also.


My MS2 turnoff is more bendy, something that can be a problem for a switching area. Brourd’s MS2 turnoff is almost unnatural straight in the second state. However I would love to see this straight MS2 turnoff put onto other designs, to see if it still gives raise to extra high fold changes.


I think we have a very good shot at moving elements onto other designs as long as they are not supposed to directly interact with the new design they are moved into. A MS2 and its turnoff can regularly be individual self sufficient systems as long as they are in 1 and 2 input systems.

Sensor B MS2 OFF

Main type among the winners.

  • B before MS2

  • B at bad spot (5’ side) - easier for kickoff

  • Design complex at 5’ end position

Sensor B MS2 OFFjpg

As the Sensor B MS2 ON lab, this lab with the strong B sequence shows bigger tolerance than usual.

A good deal of the winning designs have very bending binds between design and sequence B.

JL mod by mat, score 94%


Just like the Sensor B MS2 ON lab, this lab has high fold change error rates. And also some surprisingly low cluster counts in one of the states.

Now I wonder if there is a relation between the designs having these strong bending pairing between sequence B and the RNA design. And there really seem to be. The top 2-5 has low cluster counts and high fold change error rates. And these designs are siblings of the above shown design.

Sensor B MS2 ON

This lab has a tendency for reversal of the sequence landing spots. The B sequence is a rather strong one. And as such I suspect this effect the equation. The strong sequence has the ability to also hook up at a less optimal place - early in the sequence, where the much weaker A sequence is more dependent on getting the optimal spot late in the RNA design.

Notice that Lroppy’s winner do not have the B sequence bound up at the furthest 3’ end - which is generally the best site to place a sequence one wants to catch and turn on. Because this sequence is strong, so it doesn’t need it.

Sensor B MS2 ON winnerjpg


Design by Lroppy, score 98%

Other winning variant, of more like the Sensor A MS2 ON pattern.

Sensor B MS2 ON winner2jpg


So this lab has high scoring designs that do not all follow the same path of the Sensor A MS2 on lab as neatly as the design above. Where MS2 would have been followed by the B sequence, last in the design at the 3’ end.) Instead some of the following patterns pops up in high scorers. (Along with more)

  • Reversal of the sequence binding spot, so sequence B comes before MS2

  • A static stem in between MS2 and B. (With MS2 first and sequence B last, although also sometimes reversed)

  • The solve can pop up further towards the 5’ end - which is generally worse. (Except for the miRNA-in, reporter-out lab, round 1 - for whatever reason. The only thing that makes it stand out from the round 2, is that it has a real short oligo involved.)

Though there are more of these among the designs winner designs with reversed binding site pattern that have a higher fold change error rate. As I should have mentioned earlier, I’m generally using a cut off line in the data on a Fold change error rate of 1.15. I have been using this cutoff line lately. Most of the data is good enough to leave enough winners back afterwards.

This lab in particular has a higher fold change error rate than the others. Not sure why.

[A]/[B] with predefined binding site

The high scorers follows this pattern.

  • B before A

  • MS2 turns up at end position

Malcolm's JR mod

Predefined binding sitejpg

Design example:

The related but unconstrained lab (R2) has its the MS2 positioned in the middle in the winning designs.

In comparison the lab with predefined binding site forces the MS2 to shift place to the end of the sequence. No design scoring over 73%, have their MS2 placed in between the oligos.

Thereby the lab looses its option of having the MS2 and the B sequence sharing lane (or really rather magnet segment) as was the case in the majority of the R2 winners. The constrained lab did achieve a similar scoring winner. But it seems it was a bit harder making winners in the constrained lab. Time will tell.

However it was yet again confirmed that MS2 can be happy at ends of the sequence - at least as long as there two sequence inputs. Having MS2 at ends were generally less optimal in single input labs and Riboswitch on a chip labs. There it generally helped having MS2 held from both ends to have a chance to get it turned off. But when MS2 are having similar length playmates, it becomes more willing to take the odd position and still fold open and be turned off - even if not being held from both ends.

[A]/[B] with predefined binding site (Alternative)

  • A before B

  • MS2 at middle position

Salish mod of JR

Predefined binding site alternativejpg


Design example:

This is absolutely opposite to what would happen had the binding sites not been predefined. The weaker TB A gets the worst spot for turn on. See the R2 overview:


While Salish solve shows it is possible solving the other way too. The predefined lab, forces reversal of the pattern that generally pops up in R2 winners. But it seems to be just as solvable this way. The global folds are even better for the predefined lab.

It seems that when two inputs are involved, they are a bit less picky about their landing spot, compared to when single inputs are involved.

I think when two input gets involved and one have a MS2 in the middle and the oligos far apart, the sequences put in between gets to play a stronger role, one that also allows for reversal if things are balanced right.

I however still think one particular way of solving will be better than the other. And as for now I’m betting on the unconstrained one.
Photo of Eli Fisker

Eli Fisker

  • 2229 Posts
  • 490 Reply Likes


Single input rules

Want a single input lab turned on? - Place the sequence at the best spot (3’ side) in state 2. Place MS2 before. (Example labs: Sensor A MS2 on, microRNA 208a)

Want a single input lab turned off? - Place the sequence at the bad spot (5’ side) in state 1. Place MS2 after. (Example lab: Sensor A MS2 off, Sensor B MS2 off)

Alternative turnoff route is to give the MS2 a turnoff sequence and split the pairing of the oligo on both sides of the MS2. (Sensor v3 - variant 1 and variant 2)

Sum up: The late bit in the RNA sequence (3’ end) is best for whatever one wants turned on. The early bit in the RNA sequence (5’ end) is best for whatever one wants turned off.

Only if the oligo is real strong like the B one, it can shift balance and side. Then it can sometimes take the bad spot or less optimal spots and still get turned on.

Double input rules - Exclusion labs

If possible, let the sequences overlap/share lanes. This is a quick and dirty way of achieving an exclusion switch, where one sequence kicks out the other. If not possible, place the MS2 in the middle and the oligos to the sides.

Place leaving sequence at the bad spot (3’) end and the staying sequence at the best spot at 5’ end. (Which doesn’t have to be exactly at the 3’ end in the RNA design always - it's just the oligo you want to stay that has to be 3’ in relation to the oligo you want kicked out.)

(Example labs: R2, R3, miRNA-in, reporter-out round 1 and 2)

Basically this is just the two options in the single input rules above, added together.

However as a twist the predefined labs showed that other patterns would pop up, that were working too, when binding site was forced. In the alternative binding site lab, the reversal of the above mentioned pattern seemed to work just as well. But I suspect that one will turn out better than the other.

Exclusion labs without MS2

If possible, let the sequences overlap/share lanes. Put the oligo you want turned off before, or 5’ end of the oligo you want turned on.

It seems the miRNA-in, reporter-out with short reporter, round 1, preferred to be at the bad spot at the 5’ end while the miRNA-in, reporter-out with long reporter, round 2, preferred the better 3’ spot. (Example labs miRNA-in, reporter-out round 1 and 2). Something I don’t get why yet.

What sequence strength may reveal

A thing I count worth looking out for.

The sequence B is really strong - it has lots of strong bases.

I already noticed that the single input B sequence labs alone got skewed from the regular pattern that pops up in the pure Sequence A labs and that I was expecting. I think we can actually use this for something.

I strongly suspect that we can use the strong sequence B to investigate which crazy solves each lab will tolerate - also for the A and B mixed labs - where the weaker sequence A when alone, will highlight what areas are bad for a solve, by gleaming by its absence and extremely bad scores.

This Sensor B MS2 ON lab had a lot of crazy possible solves among the near winners.

Photo of Rocketdog42


  • 33 Posts
  • 2 Reply Likes
Thank you, Eli, for all this analysis. Probably will take me a month to absorb it!
Photo of Omei Turnbull

Omei Turnbull, Player Developer

  • 977 Posts
  • 305 Reply Likes
@johana: I was looking at the switch graphs and realized that the target structure is the same for every puzzle.  For example, here's the target structure for a Sensor B MS2 ON submission, which doesn't make sense for that puzzle.

Photo of johana

johana, Researcher

  • 96 Posts
  • 45 Reply Likes
@omei: Yes, for the sensor puzzles there are only two states, so the target is simply yellow or blue. The heatmap refers to all the conditions we tested, including those that were used for other puzzles. It gives us a sense for how well the design behaves compared to the final goal of finding a good A/B calculator.
Photo of Omei Turnbull

Omei Turnbull, Player Developer

  • 969 Posts
  • 304 Reply Likes
Thanks.  I think it will cause some player confusion about how to interpret the Target/Result graphics, when a design getting a score of 100 has Target and Result graphics that don't look at all alike.  But I guess we'll just have to try to get the word out that the scoring is customized for the goals set for each puzzles, while the Target graphic reflects the A/B target, not the target of the submissions' puzzle. 
Photo of Jennifer Pearl

Jennifer Pearl

  • 203 Posts
  • 31 Reply Likes
Ran R100 through DPAT and got data on the designs that scored 80-100 and the 30 60 groups. I posted it at a strategy market here:

( I messed up the round number in the title sorry)

This data can be used to assist in designing for the next round. It contains report files on the secondary structures like how many of the same structures are used as well as markers for those score ranges. I also included the raw DPAT data.

I was thinking. If you use the marker data use #Markers as one of your hashtags please and #PartitionFunction if you use the partition function values, and #PairingProbs if you use pairing prob data..
Photo of Eli Fisker

Eli Fisker

  • 2229 Posts
  • 490 Reply Likes
RNA as a Teeter tooter

What I wish to convey to you, is the image of shifting balance.

Equal sized playmates and the game can begin. Each taking terms of being up and down.

Teeter-Totter Playground Children Game Kids


In our switch RNA designs, the teeter totter is unevenly adjusted from the start.

Not each seat is equally hospitable. Not each playmate being equally strong. One sequence has much weaker base content than the other. To achieve balance and a good game, things needs to be adjusted.

Balanced seesaw between playmates. Both are going to get an enjoyable ride.

Teeter-Totter Playground Children Game Kids


When the bigger sibling moves toward the middle, balance can be achieved and the smaller sibling will not get a too bumpy or flying ride. Now most seesaws are not adjustable like the one above.

TB A sequence with weak bases is the little brother sequence, needing the better spot on the teeter tooter.

The teeter totter being the RNA sequence. The oligos being the kids. MS2, a reporter sequence or a static stem can be the wooden stem.

In case of two RNA inputs, MS2 will tend to take the middle spot, unless things needs to get really skewed - like in the OR lab. What to place where, will also depend on what you want done.

Single and double input labs

As I mentioned earlier:

The late bit in the RNA sequence (3’ end) is best for whatever one wants turned on. The early bit in the RNA sequence (5’ end) is best for whatever one wants turned off.

While this happens for single input labs, this seems to be what is naturally inclined to happen, even in double input switches where one sequence needs to leave. This be both if they are sharing lanes or being apart.

Getting the balance right

However it is very well possible reversing the order of the oligos and place the weak oligo at the seemingly worse spot. As was done in the one lab with predefined binding side: See the section: [A]/[B] with predefined binding site (Alternative)

It just takes adjustments of some of the other levers, to shift balance between the sequences and the binding spots. Just to mention a few.

  • Strength of the individual RNA sequence matter

  • Which end the design complex is placed, matter

  • What end the oligo is placed at, matter

  • A complement prolongment can shift balance

  • A MS2 turnoff sequence can change balance - short distance turnoff

  • Which side a MS2 is turned of at, can shift balance

  • A tail magnet can shift balance - long distance turnoff

  • A static stem in the switching area can shift balance

  • The placement of MS2 inside the RNA sequence matter - not equally effective at ends

  • And so on

Driving forces

  • The concentration of the oligos are the switch motor, their concentration a driving force for switching.

  • Oligo strength and concentration will determine where they will be most likely succeeding in landing and most likely leaving

  • Extra sequence in between switching elements in for turnoffs

  • Lane sharing - oligo concentration either welcome or force oligo out

  • MS2 used as turnoff for oligos

Example of shifting balance

I have shared an example of shifting balance. In the R2 lab it was not possible for me to make the sequences overlap at the order I wanted them in - the leaving sequence before the staying sequence and still have them share lanes. Instead I did something else to shift the balance.

Related post: Grammar of the RNA
Photo of Eli Fisker

Eli Fisker

  • 2229 Posts
  • 490 Reply Likes
The position of the design complex

I have been wondering about what decides at what end the design complex turns up. I have generally wanted to put them at 3’ end, after seeing the initial success from the microRNA 208a lab.

However this turned out to be not the full truth. I think there is something else that determines what end the design complex ends at.

The A/B designs with A and B sequence turnoff, tended to have their design complex at the ‘5 end of the sequence instead.

Basically the single input ON and OFF labs, the ON labs have had the design complex at the 3’ end and the OFF designs have had the design complex at the ‘5 end. Each choosing the position that sent the MS2 towards the middle of the design.

I think the R3 winners were moved towards the 5’ end, because the design complex had its MS2 last. Had the design complex been moved to the 3’ end, the MS2 would have been put at the absolute end of the design, which it has generally shown to work worse - although it is more possible in designs with two inputs compared to one - something which is also shown in the [A]/[B] with predefined binding site lab.

If MS2 is at one end in the design complex and not between oligos, then the design complex seems to choose whatever side that gives the MS2 a more middle position over an end position.

The pattern with the design complex being skewed to one end of the design, typically happens when sequences can be made share lane. This result in much smaller area of the design, where designs that can not be made sharing their switching elements, tend to take up far more of the available space.

Lab quest: Which end is the best position in the RNA design

Anyway I suggest and am planning myself to moving some of last round successful designs in open labs, to the opposite end of the RNA sequence to see if they have any particular bias in relation to placement - which I strongly suspect they will.

I mark these with the hashtag: #Design complex moved to 5’ end or #Design complex moved to 3’ end.

Oddball case

Especially the winners in the two miRNA-in, reporter-out labs, round 1 and 2 have been driving me nuts. Each has its winners turn up at either end of the RNA sequence. What I can't understand is why they don't both turn up at the same end.

Majority winner type from each lab:

Now these labs don’t have MS2’s. It can be that they are not so fussy about their placements, as have they had a MS2.

The only real difference between these labs, are that the reporter (bound in state 1) is short in round 1, but longer in round 2. Ok, the TB signatures are different also. But not by much. They are in same mod of colors/bases.

In the Single input A and B labs, the position of the MS2 seemed to be involved in determining what end the design complex ended at. Can it be that the different reporters - in MS2's place - are whats destining these design complex to each their end of the RNA sequence?

The only thing I can see they accomplish by their exact position of the design complex is to get their strongest base section put in place for a good dangling end position.
Photo of Eli Fisker

Eli Fisker

  • 2229 Posts
  • 490 Reply Likes
I meant to place this post on Global fold change and switch simplicity in this discussion but for what ever reason I put it up in another discussion instead. So to fix the mishap I add a link here.
Photo of Eli Fisker

Eli Fisker

  • 2229 Posts
  • 490 Reply Likes
R3 experiments

I did experiments to first round designs with already known results, as I expected that it would matter which end a design complex were placed at. 5’ prime or 3’. The results so far seems to confirm.  

I took jandersonlee’s 3 round 1 winners that had the design complex placed at the 5’ end and slided the design complex to the 3’ end.

All slided designs resulted in worse scores (in the 60 and 70’ties).

Design results

The low scoring 3’ positioned ones have less negative kcal in both vienna and Nupack compared to the original winning 5’ positioned ones.

More negative kcal seems to be a tell tale sign for which end is the better position for the design complex, particular in those cases where the inputs share landing spot.

Background post on kcal

MS2 positioning in the sequence

Most of these switch labs prefer a somewhat middle placement in the sequence of their MS2.

  • R2 - All the winning designs have their MS2 in the middle or late middle of the sequence.

  • R3 - All the winning design have their MS2 in the middle of the sequence. A few - early middle.

  • Sensor A MS2 off - MS2 in early middle

  • Sensor A MS2 on - MS2 in middle, late middle and a few early middle.

  • Sensor B MS2 off - Early middle, middle + a few late middle. As first round started to indicate, the stronger B input has more choice where to bind up with the sequence and still get a good score.

  • Sensor B MS2 on - Middle placed, late middle + a few early middle

  • [A]/[B] with predefined binding site winners is the exception and holds the MS2 at 3’ end. However it doesn’t yield as many winners as the other labs.

  • [A]/[B] with predefined binding site (alternative) - MS2 in middle or late middle.

Positioning of inputs

The trends for positioning of the inputs that started showing in first round continues for this round. The input positions show the same pattern as I mentioned in the last section in this earlier post.

I set a minimum cluster count at 50. And looked at the sequences for designs scoring 94% and above.

Sensor A MS2 ON

Winners generally have the A input at 3’ end.

Sensor B MS2 ON

Trend towards having B input at 3’ end, but mixed pattern shows as well. The strong B input can be placed in the middle or at either end and work just fine. Just as in first round.

One of the designs from first round, tolerated a total reversal of the design complex with raised score from 96% to 100%. B input ended at the 5’ end. Which is opposite of what I would have expected had the this lab had input A instead.

Sensor A MS2 OFF

Winners generally have A input at 3’ end.

Sensor B MS2 OFF

Winners mostly has the B input placed at 5’ end.

Input position sum up

As from the above summary, the OFF designs tend to follow same structure pattern for placement of inputs and so do the ON designs. However the different strength of the two inputs seems to be enough to throw this structure pattern a bit off sometimes.

So seemingly something as simple as variation of the input sequence and its strength opens up for different positioning of the input in relation to the design sequence. The stronger B input that holds more strong bases - and thus have better options for a strong pairing - can bind more places in relation to the design sequence and legally.

However on the other hand it seems much easier making winning designs with the weaker A input, no matter if it is a MS2 turnon or turnoff lab.

Most of the designs where I put the design complex at opposite end of the design, didn’t benefit. Which kind of makes sense as the MS2 has a favorite position in relation to the design - middle mostly and the input also has its favorite position too - in relation to if it is an on or off lab. Move the design complex and you touches both MS2 position and input position at once.

[A]/[B] with predefined binding site labs

I have been claiming that A should be coming before B, or rather that the sequence that is present in state 1 but needs to leave in state 2 the leaving input should be before the staying input. And I think it is particularly effective if the inputs overlap directly. However I have been on the watchout for if there should also be something to it for inputs that were placed more apart.

For background, see section: The "A before B" rule

First round results clearly showed that it was absolutely possible having the leaving sequence last and the staying first. [A]/[B] with predefined binding site (alternative)

For background post, see section: Reversed sequence pattern

While this still holds true for round 2, a more clear pattern now shows, where the lab that has the leaving input first and the staying input last, gets the most of the winners. ([A]/[B] with predefined binding site) Something I find quite interesting.

Perspective on input order

One thing else I’m interested in knowing is how much the individual sequence of the inputs and its strength has to do with the order of staying and leaving input also. Would the exact same pattern show, if input the inputs were opposite in strength?

Photo of rhiju

rhiju, Researcher

  • 403 Posts
  • 123 Reply Likes
Comment/question on input order - shouldn't it matter what the sequences of A and B are? For example if the 3' end of A has a lot of similarity with 5' end of B, those should probably bind to the same 'landing pad' on the RNA design. That would mean B before A in terms of where they land on the design.
Photo of Eli Fisker

Eli Fisker

  • 2229 Posts
  • 490 Reply Likes
Correct. And then that would be the way to do it. As to do the best with the hand you are dealt.

However I don't think the two are equal.

It seems to matter which end of the inputs in reality fit sequence wise to each other.

If for example we have a lab, where A input is to bind in state 1 and B is to be off, and B is to bind in state 2 and A is to be off.

If the 3' half end of input A is matching with the 5' end of input B, then I think this will be far more optimal than if the 5' half of input A is a match for overlap with the 3' end of input B.

What I'm saying is that I think there is an optimal end for overlap for an input that needs to be leaving (the 3' end of the input) and for an optimal end for an overlap for an input that needs to be staying (the 5' end of the input).

Not that it can't be done opposite.

I just don't think both solves types will reach the same score. It will be harder reaching the maximum potential with reversal of the inputs, even if they fit for an overlap.

However I have started to speculate about just how much the strength (amount of strong bases) of each individual input will affect things too. Besides their part complementarity. At least it seems to be affecting things like where the design complex ends up in the design sequence.
Photo of Eli Fisker

Eli Fisker

  • 2229 Posts
  • 490 Reply Likes
For a test

I'm guessing that this thing with input order of two inputs, could be tested if we have two inputs that have each their ends be able to overlap with each other which ever way.

A input capable of overlapping its 5’ with input B’s 3’
A input capable of overlapping its 3’ with input B’s 5’

Plus two different puzzle conditions. One where A input needs to leave in state 1 and B stay in state 2, and another where input B needs to stay in state 1 and leave for A in state 2.

NB. Them being complementary whichever way, would likely make them more equal in strength. Where in the A/B labs, B has more strong bases than A. Also it has them more equally distributed in the sequence.
Photo of rhiju

rhiju, Researcher

  • 403 Posts
  • 123 Reply Likes
very interesting. thanks for the clarification.

in the openTB labs (synthesizing now), we have some puzzles like A/C and B/C

can you make predictions for which orders (A before C vs. C before A) will have generally higher scores?

Perhaps even simpler, how about the cases with just one RNA in and then reporter binding?  any thoughts on order of A vs. the reporter (or B vs. the reporter)?
Photo of Eli Fisker

Eli Fisker

  • 2229 Posts
  • 490 Reply Likes

Perhaps even simpler, how about the cases with just one RNA in and then reporter binding?  any thoughts on order of A vs. the reporter (or B vs. the reporter)?

Ok, assuming you mean for the simple tb labs (riro style) and round 1.

Prediction for round 1 simple RIRO tb labs

For the lab sensor A and Sensor B the inputs didn’t fit the reporter in such a way that the leaving output could get first and the staying output could get last. (For examples of this solving style, I generally used this style for my 3 first submissions in these labs.)

They didn’t have the reporter and the input fit well for an overlap where the reporter were first and binding in state 1 and for the input to be last and binding state 2.

Only the lab with Sensor C, could the reporter get fitted with an overlap with sensor C and get the reporter (leaving) before the Input C (staying). So of the designs solved with the overlapping style, I expect the designs in the sensor C lab to do best.

When preferred ordered overlap can't be done

When there for some reason isn’t a good fit for a good overlap and in the needed order, other solving styles take over.

Another approach that I expect will do well is word change gaming between the reporter and the input. Something we also saw for the early microRNA labs with just one reporter and input. These did not use direct overlap. I suspect because MS2 is so strong that it needed something extra to help convince it to think about switching.

The word change gaming/bachelors dilemma approach do not rely on a good match for overlap between inputs, but creates complementarity by added sequence in between the input and reporter.  

Also one of the inputs can be made to fold up with itself in one state to help a switch.

Length of the reporter in comparison to the input

There were far fewer winners in round 1 of mir-in, reporter out - that had a very short reporter compared to the input, compared to the round 2 of mir-in, reporter out, where input and reporter were more even in size. That round 1 lab is more related to our tb riro designs than round 2.

In these two mentioned mir-in, reporter-out labs, in the main part of the designs, the reporter (state 1 and leaving) overlaps and goes before the input (state 2 and staying). Only exception are the minority cases of intersecting inputs. But which is something that I expect to grow more frequent for designs with more inputs. I already see it in the in silico solves.

I made a drawing in the bottom of the last image in this post.  

Order of input versus turnoff and turn on labs

For single input labs with one reporter, if the input will be before the reporter or after, will also depend on if it is an ON or OFF lab - DEC or INC.

The microRNA labs had a pattern of reversal depending on type.

For the microRNA ON lab, the input mostly was after the MS2 and only rarely shared before and after. For the microRNA OFF labs the input was mostly before or sometimes half after the MS2.

A similar pattern turned up in the A/B labs that had MS2 and 1 input.

Photo of Omei Turnbull

Omei Turnbull, Player Developer

  • 977 Posts
  • 305 Reply Likes
@rhiju: This touches on an issue I have been wondering about -- in terms of the chemistry is there any fundamental difference between "input" and "output" oligos, if we assume that the interaction comes to equilibrium?  For example, in a simple RNA in / RNA out experiment, Johan could switch the fluorescent labeling so that the "input" oligo was labeled and the "reporter" wasn't.  Would the fold change be any different?
Photo of rhiju

rhiju, Researcher

  • 403 Posts
  • 123 Reply Likes
good question! if we measure the fold ratios with no oligo and then in the limit of inifinite oligo, we'd get the fold ratios to be the same with either as the input and the other as the output. (assuming, as you said, that we reach equiliibrium) pretty neat. we can work through the thermodynamics if you want...
Photo of Omei Turnbull

Omei Turnbull, Player Developer

  • 977 Posts
  • 305 Reply Likes
Thanks!  I would like to see the thermodynamic analysis.  Even if I can't create the analysis myself, I can learn from what simplifying assumptions are made, as well as the final equations.
Photo of Jennifer Pearl

Jennifer Pearl

  • 203 Posts
  • 31 Reply Likes
I did a quick look at Sara's results for just filtering for markers only and it is not the greatest. There are many designs that score high in the bad mods lists but there are also alot in the bad. I did some investigating and I think that if I had filtered for Ensemble Diversity things would have gone much better. Here is a plot of Sensor A MS2 ON bad predictions. #2 is the results from just filtering for markers from previous round and #1 are those designes re-ran thorugh Sara doing just anlaysis filtering for Ensemble Diversity in the 1st State (if ED is greater than 15 for Same State then it is a bad design) and the results now show that had Sara initially required a ED over 15 (15 is the value i see across FMN and Oligo on labs) it would have only returned bad designs scoring under 80 with only 2 designs scoring above 80 and the lower scoring designs we not touched very much . I really think filtering for ED will filter out the majority of very high scoring designs. 

There never seams to be a single thing that eliminates a design. There are however many trends and RNA does tend to want to follow a "trend" but only insomuch as it feels like it. It follows the rule until it doesnt but generally stays true. Looking at plots of Ensemble Diversity for the FMN and the Oligo Round 1 I found that a value of 10 or less for Exclusion/Off labs and 15 or less for Same State/On labs had the vast majority of high scoring designs with maybe a couple outliers occasionally. I would plot Round 2 but the scores are not accessible through the Eterna UI yet for DPAT to crunch.
Photo of Eli Fisker

Eli Fisker

  • 2229 Posts
  • 490 Reply Likes
R2 and R3 analysis

I will go through designs solved in different styles for R2 and R3 and sum up what I think characterises them.

What causes high global fold change

Last round I mentioned that there seemed to be a relation between high global fold change and designs with direct overlapping inputs (MS2 included) and/or static stem in the switching area.

That relation seems to continue.

Static stem in the switching area + lane sharing

(The numbers in this post are found at cluster count at minimum 50, and score minimum 94%, with exception for this section)

In last round, the design with highest global fold change in R2 was my 92% static stemmed and partial input overlapping design. Note: The rerun of the original design didn’t score 92% as in round 1, but only 82% when rerun. Mat and jandersonlee managed to make winners out of it

Score 95%

R2 winnerpng


This time I could pretty much pick out the mutants of it (6348262), just by sorting the designs after high global fold change. :)

  • Global fold change range up till 10%

  • Fold change up till 12%

  • Staying input after leaving input (opposite to preferred order)

  • Static stem in the switching area

  • Partial lane sharing

  • Partial moving switch

There were however very few winners in this style. I also think this has to do with that the R2 lab can’t have the leaving input before the staying input while also overlapping well. Something which is possible in the R3 design, that despite having one more input - something that usually complicates things more, yielded far more winners. This lab also had the highest of all global fold changes - both in first round, but especially in this round.

Maximum overlap between the inputs and reporter

There was one notable exception among the designs with high global fold change that took a different road. The design 6 3 by dl2007.  I think this one holds a bunch of winners in it if mutated a bit.

Score 89%

  • Global fold change range up till 8%

  • Fold change up till 8%

  • Staying input after leaving input (opposite to wanted order)

  • Lane sharing between all inputs

  • Partial moving switch

This design follows the exact same approach as jandersonlee’s R3 winners that went for maximum overlap between the all the inputs. This design also gets a fine global fold change.

R3, score 100%, fold change 16.44, global fold change 12.71

Jl r3 with missing linespng

Image taken from this post + added a few missing pink lines.

R2 - Inputs placed apart style

The approach that won out for the R2 lab round 2 as the majority type, was distancing the inputs and going more full moving. This created a load of winners with the highest local fold change reached. Although they only reached a max global fold change of 3 for now.

This design type mainly hit through in a bunch of JR designs + mutants. Here is one with partial lane sharing. Worth noting is that by adding distance between the inputs, JR actually puts the leaving input before the staying input.

Score 100%


  • Global fold change range up till 3%

  • Fold change up till 23%

  • Leaving input before staying input (preferred order of the inputs)

  • Partial lane sharing/No lane sharing - depending on which JR design and mutant

  • Static stem in switching area in some cases/none in others

  • Full moving switch/partial moving switch

R3 - Inputs placed apart style

While most of the R3 winners were derivatives of jandersonlee’s round 1 winners, skyblue demonstrated that R3 can be solved with its inputs placed apart too.

Score 100%


  • Global fold change less than 1

  • Fold change up to 17%

  • Inputs non overlapping

  • Full moving switch

These solves gets high fold change but very low global fold change. There are very few of these among the winners.

R3 - Total input overlap - and inputs in good order

The lab with overall best global fold change was again the R3 one. Main part of the winners were of the jandersonlee type that also hit through as winners 1 round.  

  • Global fold change range up till 34%

  • Fold change up to 37%

  • Lane sharing by sequence overlap between switch elements.

  • Leaving input before staying input (preferred order)

Global fold change

Global fold change overview


For now it still seems that that specific things in a design makes it easier to achieve a high global fold change. Placed in order from more important to less important. The last option can be used to circumvent that one hasn’t gotten the inputs in preferred order.

  • Extensive overlapping between inputs

  • Inputs in preferred order

  • A static stem in the switching area

High fold change versus global fold change

I have been thinking about the relation between high local fold change and global fold change.

It is possible reaching a very high fold change with a high degree of freedom for switching (towards full moving and away from static) But I think it comes at a cost of global fold change and the switch being harder to control.

Just like for the FMN/MS2 switches, the designs that reached highest fold change, allowed an extra degree of switch freedom (only neck static but with static stem in the switching area removed) compared to the more partial moving switches (having 1 or more static stems besides the neck). But I suspect these less partial moving switches may be harder to make switch in a very precise way at a larger range of concentrations.  

The strictly overlapping R3 designs can be likened with the partial moving switches. By the extensive overlapping their degree of freedom is limited.

One see far higher fold changes in the TB switches with fewer inputs. The more inputs, the more limited fold change? Likely, because more inputs also means far fewer degrees of freedom.

Sum up

The overlapping style is sweet and effective when there is a good match between the input/reporter, but have a very picky sweet spot for making winners. Especially if the inputs are not able to overlap in their seemingly preferred order. (Leaving before staying)

There are several ways to go, to escape the seeming penalty of non preferred ordered inputs

It seems that moving the inputs apart and go more towards a bigger switching area is the way to go to reach a high fold change and lots of winners. As in JR’s R2 winners.

If the goal is high global fold change is the goal, and the inputs were not in preferred order, but it could be made work none the less, by adding a switching stem in the switching area and partial overlap,. As in my static stem in the switching area design type.

Similar dl2007’s R2 design with extensive overlapping of inputs also achieved a fine global fold change despite the inputs also being in non preferred order.

But when inputs fits each other for a good ordered overlap that will be my preferred method. For both high local and global fold change.

Photo of rhiju

rhiju, Researcher

  • 403 Posts
  • 123 Reply Likes
thanks eli. very useful. all of your points make sense:

Extensive overlapping between inputs

Inputs in preferred order

A static stem in the switching area

I can think of 'post hoc' explanations for them also, except I am getting stuck on the 'inputs in preferred order'. I don't understand why the RNA design should care which input oligo binding site is 5' vs. 3'.

Maybe something that breaks the symmetry is the position of the MS2 hairpin. Is that position also stereotyped?
Photo of Eli Fisker

Eli Fisker

  • 2229 Posts
  • 490 Reply Likes
Np Rhiju.

I'm a bit in doubt of how to understand this. Can you explain it a bit more?

"Maybe something that breaks the symmetry is the position of the MS2 hairpin. Is that position also stereotyped?"

Are you are asking what effect the MS2 have on two other overlapping inputs? I mean, if two inputs were there and could be tipped without the presence of the MS2.

I think as soon as MS2 is there it will want to somehow get involved. It is so strong that it will skew things a bit. And as such will affect things.

But the effect of one specific input wanting to be before the other, also show up in the two miRNA-in, reporter-out labs, that holds no MS2 and where the reporter itself just looks like another input.
Photo of Eli Fisker

Eli Fisker

  • 2229 Posts
  • 490 Reply Likes

I like the expression “post hoc”. I didn’t know.

“I don't understand why the RNA design should care which input oligo binding site is 5' vs. 3'.”

You taught me that RNA is chiral - that it is directional. I wonder if this has something to do with it?

Even the RNA gets made in a directional way by the polymerase. I could imagine this also affects the way RNA folds up and similar its binding patterns.

I am also thinking about something that popped up under our early microRNA labs. The microRNA seem to have preferred end (seed region) for binding up with the RNA sequence and similar the RNA sequence also seemed to have a favorite end in its sequence for catching the microRNA. (I described it here)

Here is a drawing. I’m trying to show how I think the preferred input order, leave a 3’ dangle in the RNA sequence, for 5’ pairing by the input, just as the microRNA labs seems to prefer binding.

If this round has taught me anything, it is that both ends of the design sequence are not equal. Even if the participating elements are the same.

I did a lot of moving of design complexes from one end to the other for the last A/B lab round - with few exception mainly in already high scoring designs - and most of these transfers made the design score worse.

I beforehand suspected that the better scoring part of them would go worse and that less negative kcal could be an explanation.

I am wondering if something similar could somehow explain why one input likes to go after another. Since a design complex seemingly care about which end of the design it ends up at.

If you sort by score and glance down over the kcal in this miRNA-in, reporter-out lab, you will see that the winners and high scorers generally have more negative kcal than the lower scoring designs. This is a pattern that particularly stands out in exclusion style labs with max two inputs (reporter included) and when these inputs are directly overlapping.

These are my thoughts for now. Not sure this explains the chemical why, which I get from what you write, is what you are really after.

Photo of Brourd


  • 450 Posts
  • 82 Reply Likes
Well, the chemical difference between 5' N-XXXXX 3' and 5' XXXXX-N 3' may not be significant enough to cause a preference for the association between two RNA strands (assuming parity between the concentration of G-C base pairs at certain spots in the resulting duplex), but, it's possible there may be steric hindrance or possibly some unknown entropic contribution.

One thing to remember is that the game shows the termination of the RNA strand at both the 5' and 3' ends of the sequence, however, the RNA is tethered on the 3' end.

It's also possible that it is just player preference, with players preferring one solution type over another due to the ease of solving or perhaps an unbalanced number of "good" solutions were submitted in round 1 changing the modding choices.
Photo of rhiju

rhiju, Researcher

  • 403 Posts
  • 123 Reply Likes
Right, RNA is chiral, so there may be a good explanation. I just can't think of it. On the other hand, as brourd says, we may be biased by what designs got submitted early on and then got modded.  

I hadn't thought about the free energy explanation in a while (i.e., this link that Eli gave) -- that may account for why putting design elements close to an end  is preferable. Thanks for the discussion so far...