RBS Designer — Translation Initiation Rate Prediction & 5' UTR Design

OSTIR (Roots, Lukasiewicz & Barrick, Journal of Open Source Software 2021), which continues the last open-source release of the Salis lab's RBS Calculator and re-fitted the thermodynamic model's coefficients against ViennaRNA's energy parameters. Because of that refit, OSTIR values are not interchangeable with numbers from RBS Calculator v2 — don't mix the two in one comparison. All free energies are computed by the ViennaRNA Package.

There aren't any — it's an arbitrary scale. Ratios between two predictions are the meaningful quantity ('this candidate is about 8x the one I have now'), and the absolute number is not a protein concentration and can't be converted into one. That is a property of the model, not a limitation of this implementation.

There are two different numbers here and the difference matters. The figures usually quoted for OSTIR — 53% of measurements within 2-fold and 91% within 10-fold — are calibration residuals: they say how closely the fitted model reproduces the 132 measurements of Salis et al. 2009 that its own coefficients were derived from, which is not the same as how it performs on a sequence it has never seen. The number for that is a rank correlation of Spearman ρ = 0.39 against measured expression on two 5' UTR datasets it was not fitted to (Gilliot & Gorochowski, Nucleic Acids Res 2024). So: use it to rank candidates within one construct context, treat two candidates that score close together as indistinguishable, and don't read the absolute value as an expression level. We also measured what that rank correlation means for an actual choice between two candidates, which is the thing you are using it for: over fourteen published libraries, hosts and reporters spanning six bacterial species — and stated against the ratio of the two PREDICTED rates, which is the thing you can actually see — the stronger of a pair is ranked first 57% of the time when the two predictions are within 2-fold of each other, 72% at 3–5-fold, and 87% above 100-fold, where 50% is a coin toss. Below 1.2-fold apart the prediction carries essentially no information (53%). Individual systems vary widely around those means, so read them as a guide to when a gap stops being trustworthy rather than as a per-prediction probability. A series of designs sitting within about threefold of one another is therefore not reliably rankable — build it and measure it.

By combining a spread of Shine-Dalgarno cores (from full complementarity to the anti-SD down to minimal) with SD-to-start spacings across the biologically relevant range, then scoring every combination with OSTIR in the context of your actual CDS. No strength is asserted for any candidate sequence in advance — the ranking comes entirely from the model. The spacer is poly-A by construction so that varying spacing doesn't also introduce new secondary structure; check the returned sequence if your cloning strategy needs a particular site inside the UTR.

The standby-site term depends on the sequence upstream of the RBS, so the model needs some 5' context. If you don't supply one, a 20 nt unstructured poly-A leader is assumed and the result says so. For a construct-specific number, paste the real transcribed leader your promoter produces.

Partially. The model is parameterized on E. coli, but you can supply a different anti-Shine-Dalgarno sequence (the 16S rRNA 3' end) for your host, which is the single largest host-specific term. Everything else in the model stays E. coli-derived, so treat a non-E. coli prediction as a ranking heuristic rather than a calibrated rate.

It models translation initiation only. It says nothing about transcription, elongation, mRNA stability, protein folding or toxicity, codon usage, or whether the designed UTR introduces a restriction site, cryptic promoter or out-of-frame upstream start codon in your final construct — check the returned sequence for those separately. The Construct QC Linter covers several of them.

Yes — both rbs_predict and rbs_design are callable from the REST API and the MCP server. Because each run does multi-second RNA folding on a shared service, they are rate limited; the error tells you how long to wait if you hit it.