From a target to a ranked oligo.
Naming a disease gene is the easy part. The program is decided by the molecule: which site is even reachable on a folded transcript, whether to silence it, splice-correct it, or base-repair it, how to discriminate a single pathogenic base from the wild-type allele, and how to stay off every paralog in the transcriptome. Coracle Oligo takes a target and an intervention and returns a ranked, gated set of oligo designs, each carrying its drive decomposition and a stated validation tier. Deterministic. CPU. No per-target training.
These are real engine outputs on documented targets. Coracle Oligo is in preview; to run your own target, request access below.
Three interventions, one engine.
Gene silencing, splice correction, and ADAR base-repair share one design funnel. The readouts are reproduced against the literature and against real editing data, and each result is tagged with what it earned: validated where a benchmark holds, calibrated where a measured envelope anchors it. The full readout list is under Capabilities; a few headline results are here.
What the engine returns.
The pruned, ranked design map
design(target, verb) returns the gated, ranked set of oligo designs for
a target and an intervention, each with its full drive decomposition and the gates it
passed. The interventions are gene silencing, splice correction, and ADAR base-repair.
This is the deliverable: candidate oligos with their reasons, ordered, with the dead
sites already removed.
Allele selectivity
The single-base discrimination between a pathogenic allele and the wild-type, from one mismatch and a dosing window. Reproduces the field's reported 5×, 124×, and beyond-100× selectivities from a single mismatch energy. A discrimination ceiling and the dose that reaches it; the in-cell knockdown percent stays out of scope.
Target accessibility
The opening free energy of a candidate site on the folded transcript, the map of open, bindable windows. A buried site is rejected before anything downstream runs. Computed from the transcript's own folding ensemble.
Off-target landscape
A seed-and-extend scan of the candidate against the transcriptome, returning the worst credible off-target hybridization and a specificity margin. Scales to a real transcriptome. The margin is a relative measure on one consistent scale.
ADAR base-repair design
For a pathogenic adenosine, the on-target-versus-bystander guide design: where to place the editing base, which neighboring adenosines to protect, and the editability and bystander-specificity of the result. The neighbor-rule benchmark holds at ROC-AUC 0.747 against real human editing contexts (shuffle p = 0.005), and the engine reproduces a worked SERPINA1 PiZ-class design at bystander-specificity 0.81. A design ranking, not an in-cell editing yield.
Gene silencing
siRNA guides ranked by a fitted efficacy readout (gene-grouped cross-validation Spearman 0.63, with no per-target training and no leakage across genes), and RNase-H gapmer designs ranked by binding stability. The two mechanisms are kept separate, so a guide rule never scores a gapmer. A ranking signal that transfers across assays.
Splice correction
Splice-switching oligos that mask a regulatory element to shift isoform choice, scored with a standard splice-site model and the element coverage. Reproduces the nusinersen mechanism on the real SMN2 exon-7 / ISS-N1 silencer, sharing a 12-nt core with the approved oligo. A mechanism reproduction on a documented target.
Chemistry-aware thermodynamics
On-target stability computed with the correct backbone per chemistry: DNA/DNA, RNA/RNA, and RNA/DNA hybrid for a gapmer's gap, plus per-substitution increments for the LNA, 2'-MOE, and cEt wing modifications. The hybrid table reproduces its published worked values to two decimals. Each imported scale is marked and sourced.
Deterministic, sub-second, CPU
The design funnel is deterministic: no GPU, no queue, byte-identical on re-run. The same target returns the same designs a year from now. It runs identically on natural and designed sequences, including private mutations that no model was trained on.
Refuses outside competence
Every output states what it does not claim, and the engine declines targets outside its tested range rather than returning a confident-looking number. Honesty is the feature.
Honest by construction.
Every output says what it is load-bearing on.
Each design carries a validation tier, an estimate kind, an explicit statement of what it does not claim, and a provenance record (model version, schema version, request id). Designs are deterministic and reproducible.
On commercial terms, designs for your private targets remain yours. The engine does not train on customer data, because the design engine does not train.
{
"verb": "splice_correct",
"validation_tier": "validated",
"estimate_kind": "element_masking_design",
"scope_note": "design ranking, not a clinical outcome…",
"provenance": {
"model_version": "0.3.0",
"request_id": "c98e25c1…"
}
}
In preview. Open for collaboration.
The protein and RNA engines are live at protein.coracleresearch.com and rna.coracleresearch.com. The oligo engine is in preview. A short email gets you a walkthrough, a design run on your own target, or early access for a program.