Fable 5.2 predicts protein contacts from sequence alone — but runs don't reproduce
sokrypton · x · 2026-11-19
sokrypton shares an empirical analysis of Fable 5.2 on protein contact prediction from sequence alone (Cβ within 8Å, |i-j|≥6):
- Results were not reproducible run-to-run with the same simple prompt; within the session that worked, he had the model reconstruct its method into a detailed prompt, published as a GitHub gist.
- The attached figure shows contact maps for 8 test proteins (ubiquitin 69% correct, TEM-1 β-lactamase 30%, Top7 65%, L29-like only 19%), with blue/red dots marking correct/wrong predictions by confidence.
More from Models
- GPT-6.2 Astra users note it can work 30 minutes without any progress updates — Angaisb_ · 2026-11-19
- LMArena launches six new agent indexes; Claude Fable 5.2 tops Work, Coding and Chat — arena · 2026-11-19
- Investor complains of 'rude' Anthropic rate limit, blames Nvidia deal monopoly — StewartalsopIII · 2026-11-19
- Hands-on with GPT 6.2 astra and sol: not an Opus 5.5 drop-in, but 'really good' once you adapt — alexcovo_eth · 2026-11-19
- 3x-repeat benchmark: Fable 5.2 tops Opus 5.5 on most protein contact prediction tasks — sokrypton · 2026-11-19
- LMArena overhauls agent leaderboard: Agent Arena Index scores 2.9M sessions with anchored 1000-point scale — arena · 2026-11-19