De Novo DNA alternative: the Salis lab calculators, without an account

OSTIR, with free energies from the ViennaRNA package — an open-source implementation of the thermodynamic model of translation initiation, so the whole calculation is inspectable rather than a black box. The response carries the full breakdown: the 16S rRNA hybridisation term, the cost of unfolding the mRNA structure over the site, the SD-to-start spacing penalty, the standby site and start-codon binding. It also carries what the model is worth on data it was not fitted to — a rank correlation of about 0.39 with measured expression — because a rate with no accuracy beside it invites you to plan a yield around a number that cannot support one. Treat the output as a ranking of candidates within one construct, which is what it is good at.

Three reasons. You want an answer now rather than an account — nobody creates a login to check whether their 5' UTR folds over the start codon. You want the calculation from a script or an AI agent instead of a web form, which is what the REST API and the MCP server are for. Or you are already working in a file — a .dna, an .ab1, a GenBank — and want the analysis where the file already opens, rather than pasting sequence between two sites.

Because the published classifier's headline accuracy does not mean what it looks like, and a number that looks like an answer is worse than no number. Of its 1,076 training sequences only 303 were real orders: 373 are negative controls designed on purpose to violate vendor filters, and 400 are positive-control subsequences of fragments already known to have synthesised. The split is stratified by class rather than by source, so the held-out set is about 72% synthetic controls too — what the figure establishes is that deliberately-broken sequence separates from known-good sequence, which is not the question you have about your gene. Its labels are also one vendor's turnaround times from before 2020, and vendors have got faster since. So SeqBench measures every determinant that model uses — repeats, GC extremes and swings, homopolymers, hairpins — against the thresholds vendors publish, and tells you which measurement is out of range and by how much. That is the actionable half, and it does not go stale.

Yes, literally. The scaffold variants cannot be generated — a Cas9 scaffold's secondary structure is what the protein binds, so randomising it destroys the part — they had to be measured, and Reis et al. (2019) measured 27 of them. SeqBench uses that published collection from the authors' own MIT-licensed release. What differs is the selection: SeqBench selects across promoters, handles and insulators at once, because a handle and a promoter can each be clean within their own pool while sharing 18 bp with each other, which leaves the array exactly as recombinogenic as no selection at all.

Free, with no account and no key. Sequences you paste into a tool are used to compute the answer and are not stored; the file parsers run in your browser, so .ab1, .dna and GenBank files are not uploaded at all. Sign-in exists only as somewhere to attach saved work, and nothing on the site is gated behind it.

The deterministic ones, yes, and that is most of what this suite does. Which parts share sequence, what a consensus match is, what an assembled operon carries, what base-pairs over a ribosome site — all of these are consequences of the input, and every one is pinned by tests that re-derive it independently rather than recording last week's output. The non-repetitive selection is additionally checked against the subsets the ELSA Calculator itself ships, which is the only way to catch a wrong definition rather than a wrong implementation: at the strictest threshold both reach the same count, and at a looser one this implementation keeps a strict superset of theirs.