Protein LM hits perplexity 2.2 on held-out proteins, beating ESM-class models

proteinrosh · x · 2026-11-18

Kexin Huang shares a surprising finding: Fable 5.2 shows massive gains over protein language models on next-amino-acid prediction for less-studied proteins. proteinrosh contextualizes it: even ProteinMPNN, an inverse folding model given the structure, only reaches 4-5 perplexity on truly held-out proteins — yet Fable 5.2 achieves 2.2 without structural input, raising questions about where the gains come from.

Related event: Protein Model Perplexity Debate: Does 2.2 Mean Structural Understanding?(3 posts)→

Original post →

More from Models

Models channel →