RNAValidate: CPU-only validator for AI-predicted 3D RNA structures

A predictor-agnostic linter that checks AI-predicted 3D RNA structures against experimental data (FRET, cryo-EM, SHAPE/DMS probing, designability) — and flags the ones that will hydrolyze before you can measure them.

The gap

Predictors and designers (AlphaFold-3, RiboSphere, RoseTTAFold-RNA) are plentiful; an open-source, reproducible layer that audits a predicted 3D RNA structure against experimental evidence is not. Validating before synthesis saves thousands of USD per design.

What it does

RNAValidate reads the PDB + JSON you produce and applies five rules:

• R1 — FRET consistency: RMSD of structure-derived vs FRET distances > 8 Angstrom => FAIL.
• R2 — cryo-EM correlation: map/model CC < 0.5 => FAIL (skipped without a map).
• R3 — Chemical probing (SHAPE/DMS): >40% of probed positions contradict the structure => FAIL (skipped without data).
• R4 — Designability: too few residues / broken backbone / no foldable motif => FAIL.
• F1 — Hydrolysis propensity / tube false-positive (DERIVED): score > 0.60 => likely to degrade before measurement.

Transparency note

F1 is a DERIVED rule (risk R5), proposed by the author and not extracted from a single paper. It needs experimental validation against measured degradation (tube half-life) before promotion to a hard rule. Thresholds (FRET RMSD, cryo-EM CC) are documented chemistry, not tuned to the fixtures. Fixtures use a minimal synthetic hairpin PDB plus example FRET (external-validity, not circular).

Results

• 35 tests passing, 96.84% coverage (gate 80%)
• CPU-only, 0 external predictor invocations (AC-8)
• AGPL-3.0-or-later

Try it

pip install -e .
rnvalidate check --in structure.pdb --exp exp.json --format json

Stack

• Python 3.11+, Typer
• pytest + pytest-cov for the suite

Links

• Repo: https://github.com/amurlaniakea/rnvalidate

---

License: AGPL-3.0-or-later (c) 2026 Pedro Sordo Martinez — [e-mail]