FULL STORY
Claude's Surprising Protein Structure Predictions Put to the Test
After users noticed Claude Fable 5.2 could predict protein sequences, researchers including sokrypton ran experiments showing the model can partially predict contact maps and secondary structures, sparking wide community discussion.
2026-11-18 ~ 2026-11-19 · 2 episodes · 14 posts
Episode 1 · sokrypton tests whether Claude Fable can predict protein contact maps from sequence alone (2026-11-18, 12 posts)
On November 18, researcher sokrypton posted a series of experiments showing that without invoking structure prediction tools like AlphaFold, the protein language model Claude Fable 5.2 could predict residue contact maps in folded structures from amino acid sequence alone (residue pairs with C-beta atom distances under 8Å and sequence separation ≥6). He first asked the model to return the 128 highest-confidence contact pairs for a 68-residue de novo designed protein; the predictions partially matched the true contact map. He then ran a 286-residue sequence, having the model return the top 512 pairs ranked by confidence, and shared a side-by-side comparison of Fable's output with ground truth, concluding that "parts of it look correct."
Confirmed
- sokrypton demonstrated several experiments: a 68-residue de novo protein (returning 128 contact pairs), a 286-residue sequence (returning 512 contact pairs), and a prompt combining dynamic programming and linear algebra that had the model reason out the folded structure of ubiquitin from sequence alone, with partially correct predictions
- The motivation was to use an LLM to test whether the model genuinely understands proteins rather than relying on structure prediction pipelines
- Because extracting logits requires multiple runs per position and is costly (this week's token budget was exhausted), sokrypton proposed switching to an alanine scan to approximately test the model's response to protein mutations
Unconfirmed
- proteinrosh suggested using a Jacobian method to estimate how large models respond to mutation effects, to determine whether they understand long-range protein contacts—but this remains an unimplemented idea; whether Claude's success reflects reasoning or memorization is still unresolved
- proteinrosh also offered a counterintuitive take: memorization isn't necessarily bad—language and reasoning capabilities can extract information from representations in different ways, and a model memorizing vast numbers of sequences isn't inherently harmful—though this is a personal opinion, not a conclusion
Why it matters
- If LLMs can genuinely "reason" out structural information from sequence alone, this challenges the default assumption that structure prediction must rely on specialized pipelines like AlphaFold, and offers a lightweight path for probing whether large models truly understand biology
- Jacobian-based mutation analysis could test if LLMs grasp protein contacts — proteinrosh · 2026-11-18
- Researcher argues memorization isn't inherently bad; Jacobian method could test LLM reasoning — proteinrosh · 2026-11-18
- Approximating Claude's protein mutation response with an alanine scan — sokrypton · 2026-11-18
- Researcher shows Fable predicting protein contact maps straight from sequence, top 512 pairs ranked — sokrypton · 2026-11-18
- Fable's predicted protein contact map vs ground truth: partly correct — sokrypton · 2026-11-18
- Claude Fable 5.2 correctly predicts de novo protein contacts from sequence alone — sokrypton · 2026-11-18
- A single prompt gets Claude to predict contacts of a de novo protein using DP and linear algebra — sokrypton · 2026-11-18
- LLM predicts ubiquitin's protein fold from sequence alone via a clever contact-map prompt — sokrypton · 2026-11-18
- Community test: Claude Fable 5.2 predicts protein contact maps from sequence — LeoTZ03 · 2026-11-18
- Testing Claude Fable at direct protein 3D coordinate prediction: TM-scores of just 0.19-0.25 — LeoTZ03 · 2026-11-18
- sokrypton predicts protein structures from LLM contact maps, admits the structure part is a gimmick — sokrypton · 2026-11-18
- sokrypton shows models can reason about which beta sheets interact — sokrypton · 2026-11-18
Episode 2 · Claude Reportedly Predicts Protein Structures Natively (2026-11-18, 2 posts)
Tests claim Claude Fable 5.2 can natively predict protein structures such as ubiquitin folding without specialized tools, with follow-up experiments designed and analyzed by Opus 5.5 drawing wide attention.
- Claude correctly folds ubiquitin from a plain prompt, no specialized tools needed — iScienceLuvr · 2026-11-18
- Claims that "Claude Fable 5.2" predicts protein structures, verified by Opus 5.5 — anshulkundaje · 2026-11-19