EF-Synonymous · synonymous variant scoring

Input

Search by gene (Ensembl REST) — optional

Transcript CDS sequence (A/C/G/T only, no header)

Paste the full CDS (or a fragment with ≥64 nt on each side of the variant). Position 1 = first base of the start codon.

HGVS variant (c.)

Format c.{pos}{REF}>{ALT} at CDS level.

Result

–

P(pathogenic) · shaded gray zone = no evidence · scale 0 (benign) → 1 (pathogenic)

ACMG evidence: –

Glass-box breakdown

The prediction is the sum of two interpretable physical contributions (logit). Nothing opaque.

ΔG stacking profile (Turner) around the variant

Φ native (WT) Φ mutant ±10 nt window (σ) vertical line = variant position

SaMD traceability: features, contributions, model version and CDS fingerprint

How it works (and what it is not)

The model is a logistic regression over two physical observables matching the dedicated CNN of Paper 1 (AUC 0.683):

σ (sigma_signed): signed sum of nearest-neighbor stacking ΔΔG (Turner RNA parameters) over ±10 nt around the variant. Measures the local thermodynamic destabilization of the mRNA.
is_GtoA: G>A transition (CpG mutational bias, 2.60× enriched in pathogenic).

All inference runs in your browser with numpy ported to JavaScript — no server, your sequence is never sent anywhere. This is the glass-box argument for SaMD/ACMG: every prediction is reproducible, auditable and traceable variant by variant.

Honest limitations: AUC ~0.68 (supporting evidence, not a verdict). Wide gray zone. Illustrative ACMG mapping thresholds, require local calibration. Trained on ClinVar synonymous variants (v2, n=1957). Do not use for clinical decisions.