What a design report cannot tell you
A design report is a ranked list with reasons, not a prediction of experimental success. Six things it cannot tell you, why each one is out of reach, and what to do instead.
Software that ranks sequences is easy to over-read. A table sorted by score looks like a verdict. It is not, and the gap between “high score” and “works in your cells” is where most of the disappointment in this field comes from.
Here is what a design report cannot tell you, stated plainly.
1. Whether your construct will express
No sequence-based model predicts protein yield with useful accuracy. Expression depends on transfection efficiency, cell type, transcript stability, codon-tRNA pool matching, structure around the start codon, innate sensing of the RNA itself, and the assay you use to measure it. A design tool sees the sequence. It does not see the rest.
What it can tell you: this candidate has a clean 5′ UTR, no structure over the AUG, GC within the vendor window, and no homopolymer runs. Those are necessary-ish conditions, not sufficient ones.
2. How many genomic off-targets a guide has
A real off-target count requires aligning your spacer plus PAM against a reference genome and scoring mismatch tolerance at each position. That is a different computation from anything that runs in a design step: a multi-gigabyte index, minutes of compute, and a scoring model.
Our report gives you an internal seed-match count — useful for spotting repeats in the region you supplied, and nothing more. When a tool reports “0 off-targets” for a genome without saying which genome and which scoring, treat the number as marketing.
3. Whether a molecule is patentable or free to operate
Design output is a sequence. Freedom to operate is a legal question about claims that may cover that sequence, its modifications, its delivery, or its use. No amount of computation answers it. If your programme is commercial, this is a search and an opinion letter, not a design step.
4. Whether your institution will approve the work
Biosafety and biosecurity review is institutional. A tool can screen for prohibited patterns, and a synthesis vendor will screen again, but approval to work with a controlled organism or a dual-use construct is granted by a committee that reads your protocol, not your sequence.
5. Whether the folding estimate is the real structure
The ΔG in our reports comes from a greedy base-pair model: scan for complementary pairs, stack them, add a penalty per loop. It is fast, deterministic, and directionally useful for comparing candidates. It is not a thermodynamic minimum-free-energy calculation — no partition function, no coaxial stacking, no pseudoknots.
If you need a real structure, run the sequence through a proper folding package and treat our number as what it is: a ranking signal that says “this candidate has more predicted structure than that one”.
6. What the vendor will actually do with your order
Synthesis providers apply their own screening and their own manufacturing constraints, which change over time. A sequence that passed our QC panel can still be refused, delayed, or manufactured with a note. The vendor’s decision is the vendor’s.
So what is the report for?
Three things, all of which are real value:
Elimination. Most bad candidates are bad for boring reasons: GC out of window, homopolymer runs, cryptic splice sites, poly(T) in a U6 context, restriction sites. Catching those before ordering saves a week per rejection.
Comparison. Given twelve candidates that all pass elimination, the report tells you which have cleaner 5′ ends, better codon adaptation, fewer predicted structures. That ordering is genuinely informative even though it is not a prediction.
Traceability. A run stores its parameters — codon table version, seed, GC window, PAM. Six months later, when a reviewer asks why candidate #4 was chosen, the answer is in the record rather than in someone’s memory.
How to read a score
A composite score of 87 versus 84 does not mean the first one is 3.5% better. It means the first one scored slightly better on the weighted combination of GC proximity, codon adaptation, predicted structure and flags. If the weighting changed, the order might change.
Treat the score as a sort key. Read the columns — GC, CAI, ΔG, flags, length — and make the decision on those. That is what a design table is for: not to decide for you, but to make the trade-offs visible before you spend money on synthesis.