LLM predicts ubiquitin's protein fold from sequence alone via a clever contact-map prompt

sokrypton · x · 2026-11-18

X user sokrypton shows that a relatively simple prompt combining dynamic programming and linear algebra lets an LLM return a protein's 3D structure from its sequence alone by predicting residue-residue contacts — and ubiquitin's fold came out reasonably correct. The prompt rules force the model to reason only from the sequence: infer secondary structure from 3/14 periodicity and amphipathic patterns, then predict C-beta contacts within 8 Å, build a 3D model from them, inspect and revise. The output visibly matches the real ubiquitin structure (1UBQ, 76 residues) rendered in 2Dmol.

Original post →

More from coding & agent

coding & agent channel →