LLM protein-folding test on fresh PDB entry lands near-random structures
MoAlQuraishi · x · 2026-11-18
- A debate over whether an LLM can fold a protein from sequence alone: @sokrypton proposed a contact-map experiment prompting the model to combine dynamic programming and linear algebra to return a structure.
- Critics note the test was toy-sized, using a small, well-studied protein.
- Rerunning the same protocol on a PDB entry released in October 2026 (MERTK ectodomain, crystal at 3.41 Å), the 3D prediction came out near random: overall TM-score of 0.14 with RMSD 34.3 Å, and per-domain TM-scores of just 0.14–0.25.
- Takeaway: on genuinely novel proteins, LLMs currently fail at direct structure prediction, and earlier demos leaned on well-known benchmarks.
More from Models
- GPT-6.2 backlash is real user experience, not astroturf: dev says Astra has quirks but Sol is a solid daily driver — pvncher · 2026-11-18
- Every's founder pushes back: we re-tested the model to the wire, OpenAI gave us no heads-up — kieranklaassen · 2026-11-18
- HF trending: Thinking Machines' Inkling-2-Small, MiniMax H3-Image and Qwen3.8-27B surge together — victormustar · 2026-11-18
- US watchdog: four top AI chatbots helped plan terror attacks despite refusals — Dense-Following-9964 · 2026-11-18
- After DeepSeek's rate limits: a 4-step guide to the official paid desktop path — sujingshen · 2026-11-18
- blink-8M tops seven static embedding models with 55.35 MTEB and 46.70 BEIR nDCG@10 — tomaarsen · 2026-11-18