mRNA sequence designTCRshows

What this tool can and cannot do. Secondary structure is folded exactly, locally, with the Zuker algorithm on Turner 2004 parameters, and half-life is predicted with DegScore 2.1 — both checked against reference values from their respective upstream implementations. Five things need saying plainly. First, exact folding of a full-length mRNA costs O(n³) and will not run in a browser, so the ΔG curve and the half-life use 100 nt sliding-window local folding (the RNAplfold approach); long-range pairing across thousands of bases is not represented. Only the start region is folded exactly as a whole. Second, the annealing itself is still driven by sequence-level proxies (CAI, GC, uridine, CpG). Structure genuinely participates only in the final “5′ structure refinement”, and is otherwise reported faithfully on the scorecard — putting folding inside the annealing loop would need a linear-time folder, which is out of scope for this version. So the name of the “stability first” preset refers to its objective, not to a guarantee that its folding free energy is lowest. Third, the DegScore degradation rate is a sum of per-position scores and grows linearly with length, while the model was calibrated on constructs of roughly 100 nt. The headline figure here is therefore the length-independent per-nt score, with half-life converted to the model’s own 100 nt scale — useful for comparing candidates, not to be read as this mRNA’s absolute lifetime in a cell. Fourth, the MRL model is a CNN we trained ourselves on the random-5′UTR MPRA data of GEO GSE114002 (3×Conv1D, 520k parameters, int8-quantised and run in-page with no runtime dependency). The variable-length version shipped here scores R²=0.771 on held-out data, and R²=0.740 / Spearman=0.805 on natural human 5′UTRs. The same architecture, evaluated under the same protocol on the fixed-length 50-mer library used by the published CNN, gives R²=0.892 against a published ≈0.93 — which says the implementation is right. It remains a correlational model: do not read it as an absolute prediction of ribosome load. Training covered 25–100 nt, and 75 of the 225 5′UTRs in the library fall outside that range and are marked ? for extrapolation. Fifth, MRL was trained on an MPRA in a single cell line under a single condition, so it cannot see any regulation that varies with condition. The clearest case is the 5′TOP motif, carried by almost every ribosomal-protein 5′UTR, which mediates mTOR-dependent translational repression and can switch an mRNA off selectively under mTOR inhibitors or the nutrient and hypoxic stress of a tumour microenvironment — none of which appears in an MPRA. Such UTRs are flagged separately as 5′TOP; judge them against your dosing context. The downstream CDS in the training constructs was a constant eGFP, so the model reflects the UTR alone and junction effects are covered by the folding side instead. All results are for research use only; verify the intellectual-property status of a UTR sequence yourself before using it in product development.