Chemist uses AI to prove open RNA designability theorem, verified in Lean 4

rbhar90 · x · 2026-08-26

Ashutosh Jogalekar of Szilard Scientific posted an arXiv preprint, "Designability of RNA Targets with Up to Two Length-2 Helices", proving a previously open theorem: in the pseudoknot-free Watson–Crick maximum-base-pair model, RNA targets satisfying specific helix-length and obstruction-motif conditions are uniquely designable.

His key point: without AI he could not have done this — he is a chemist, not a mathematician or formal-methods expert. The proof is AI-generated, human-directed, and formally verified: GPT/Codex (5.6 Sol, Ultra) did the heavy lifting while Claude Code (Opus 5) provided adversarial review across repeated cycles, with final verification in Lean 4. The main theorem RNA.atMostTwoShortHelixDesignability depends only on the three standard axioms. The Zenodo artifact includes the frozen dependency closure, verified examples, publication checkers, source and axiom audits, and clean-rebuild evidence. AI assisted across every stage: problem framing, argument development, Lean code, counterexample searches, and manuscript drafting.

Original post →

More from AGI Musings

AGI Musings channel →