A single prompt gets Claude to predict contacts of a de novo protein using DP and linear algebra

sokrypton · x · 2026-11-18

sokrypton shows you can push LLM protein structure prediction further with a fairly simple prompt combining dynamic programming and linear algebra — Claude (Fable 5.2) correctly predicted the contacts of a de novo protein.

The screenshot reveals the full reproducible prompt: given only a 68-residue sequence, the model must skip databases, predict secondary structure (H/E/C), decide fold topology, output an explicit contact list (1.5×L pairs, |i-j|>=3), then build and iteratively revise a 3D model with code execution, rendering a C-alpha trace. A neat, copyable prompt framework for LLM-driven structure prediction.

Related event: Researcher tests whether Claude Fable can predict protein contact maps from sequence alone(12 posts)→

Original post →

More from Research

Research channel →