Four codon strategies. Every one of them improved its own axis and made something else worse.
A real run on a real payload: human erythropoietin between HBB UTRs. Depleting uridine did what it says — U from 0.203 to 0.131 across 146 changed codons — and in the same construct the count of AUGs that are not the main start went from 8 to 12, while the average unpaired probability rose from 0.46 to 0.61. Minimising GC took it from 0.603 to 0.313 and took the cryptic starts to 17, the AU-rich elements from 1 to 4. Every number here is in the result below, located to the base.
1
Say what the construct is made of
Parts, not a sequence. Anything you give two or more options for becomes an axis of the round, and the round is their product. Every part carries where it came from — the library ships no natural sequence of its own, on purpose.
5'UTR
2 options: HBB 5′UTR, and a minimal Kozak leader
NCBI NM_000518.5, 1–50 · and gccRccAUGG context (Kozak 1987)
4 strategies × 2 UTRs = 8 constructs, each 845–886 nt
Run round
2
It runs on the CPU, in seconds
queued → running → done · 11.8 s
Eight constructs of ~880 nt, each folded for its minimum free energy and its full base-pairing ensemble, then scanned. No GPU, so a round does not queue behind a structure prediction. A 1.4 kb payload across four strategies takes about three quarters of a minute.
The compute sits on EU infrastructure and your sequences stay on it. That is the default, not a paid add-on.
3
Read the result
Eight rows in the order they were generated. There is no score column, nothing is marked recommended, and the table is not sorted — because a uridine count and a cryptic start codon are not the same kind of fact, and no validated way to weigh one against the other exists. Pick any two and the compare panel prints both values and the difference, which is as far as the evidence goes.
Construct round
8 linear constructs built, varying 2 5'UTRs × 4 codon strategies. Each one folded and scanned; listed below in the order they were generated.
CPU · ViennaRNA 2.7.2 · 11.8 s
Variants
Variant
5'UTR
3'UTR
Codons
nt
MFE kcal/mol
AUP
GC
Pairs
Motifs located
v001
hbb-5utr
hbb-3utr
native
886
-283.7
0.4623
0.4786
239
20
v002
hbb-5utr
hbb-3utr
uridine_min
886
-143.9
0.6051
0.3702
173
23
v003
hbb-5utr
hbb-3utr
gc_min
886
-135.6
0.5386
0.2878
207
33
v004
hbb-5utr
hbb-3utr
sampled · seed 0
886
-226.5
0.4748
0.4153
239
19
v005
minimal-kozak
hbb-3utr
native
845
-275.9
0.4601
0.4852
237
20
v006
minimal-kozak
hbb-3utr
uridine_min
845
-130.9
0.6085
0.3716
165
23
v007
minimal-kozak
hbb-3utr
gc_min
845
-121.3
0.5446
0.2852
204
33
v008
minimal-kozak
hbb-3utr
sampled · seed 0
845
-214.3
0.4792
0.4189
219
19
Generation order. These columns are not combined into a score and the table is not ranked — AUP, GC and a motif count answer different questions and no validated weighting of them exists. An empty cell is a number that was not computed, never a zero.
v001 · 886 nt · linear
Arcs are helices of the MFE structure — one arc per run of stacked pairs, thickness is the number of pairs (239 pairs in 57 helices). Bars are the ensemble unpaired probability, binned. Marks on the lower rows are sequence matches, located; they are not confirmed sites.
AUP is the average unpaired probability (Wayment-Steele et al. 2020), reported for scale and not as a target: that work estimates ~0.4 for conventionally designed mRNA, and Leppek et al. (2022) found in-cell stability a greater driver of protein output than the in-solution stability this number bears on.
Sequence scans
every AUG, classified against the main start codon
hexamers AAUAAA, AUUAAA
AUUUA pentamers with the U fraction of the 20 nt around each, and runs of overlapping UUAUUUAWW nonamers. Classes are from Chen & Shyu (1995) and are cited, not assigned: Class I turns on "several dispersed" pentamers with no stated threshold and Class III is explicitly less well defined
Watson-Crick reverse-complement matches of >= 12 nt at least 20 nt apart, extended maximally -- the sequence-predictable part of dsRNA byproduct formation. Linear scan only: a duplex across a circular junction is not found
runs of one base of length >= 6
18 site(s) searched on the DNA (T-form) sequence, for the template rather than the RNA
No match found for: Splice consensus match, Tandem repeat.
miRNA seed site: no miRNA sites were supplied, so none were searched for. This is not a finding of zero sites
Compare
v001 against
Nothing above is a mock-up: the payload on this page is the output of the same entry point a signed-in round calls, rendered by the same component. Run one on your own construct — the form is here, and the structure-prediction example is the other half of the platform.
Signed in, two things follow from a round that a single page cannot show. A UTR you bring can be kept, so the next round starts from a list rather than from your clipboard — and it is kept with its source and a content hash, so a construct that referenced it still resolves to that exact sequence. And any variant here can be carried into another round, or have one of its elements — the leader, the window around the start codon, a circular construct’s junction — folded in 3D by the structure engines, which is the same comparison the other example shows.